2 * Kernel-based Virtual Machine driver for Linux
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
9 * Copyright (C) 2006 Qumranet, Inc.
12 * Yaniv Kamay <yaniv@qumranet.com>
13 * Avi Kivity <avi@qumranet.com>
15 * This work is licensed under the terms of the GNU GPL, version 2. See
16 * the COPYING file in the top-level directory.
22 #include "kvm_cache_regs.h"
24 #include <linux/kvm_host.h>
25 #include <linux/types.h>
26 #include <linux/string.h>
28 #include <linux/highmem.h>
29 #include <linux/module.h>
30 #include <linux/swap.h>
31 #include <linux/hugetlb.h>
32 #include <linux/compiler.h>
33 #include <linux/srcu.h>
34 #include <linux/slab.h>
37 #include <asm/cmpxchg.h>
42 * When setting this variable to true it enables Two-Dimensional-Paging
43 * where the hardware walks 2 page tables:
44 * 1. the guest-virtual to guest-physical
45 * 2. while doing 1. it walks guest-physical to host-physical
46 * If the hardware supports that we don't need to do shadow paging.
48 bool tdp_enabled
= false;
55 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
);
57 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
) {}
62 #define pgprintk(x...) do { if (dbg) printk(x); } while (0)
63 #define rmap_printk(x...) do { if (dbg) printk(x); } while (0)
67 #define pgprintk(x...) do { } while (0)
68 #define rmap_printk(x...) do { } while (0)
72 #if defined(MMU_DEBUG) || defined(AUDIT)
74 module_param(dbg
, bool, 0644);
77 static int oos_shadow
= 1;
78 module_param(oos_shadow
, bool, 0644);
81 #define ASSERT(x) do { } while (0)
85 printk(KERN_WARNING "assertion failed %s:%d: %s\n", \
86 __FILE__, __LINE__, #x); \
90 #define PT_FIRST_AVAIL_BITS_SHIFT 9
91 #define PT64_SECOND_AVAIL_BITS_SHIFT 52
93 #define VALID_PAGE(x) ((x) != INVALID_PAGE)
95 #define PT64_LEVEL_BITS 9
97 #define PT64_LEVEL_SHIFT(level) \
98 (PAGE_SHIFT + (level - 1) * PT64_LEVEL_BITS)
100 #define PT64_LEVEL_MASK(level) \
101 (((1ULL << PT64_LEVEL_BITS) - 1) << PT64_LEVEL_SHIFT(level))
103 #define PT64_INDEX(address, level)\
104 (((address) >> PT64_LEVEL_SHIFT(level)) & ((1 << PT64_LEVEL_BITS) - 1))
107 #define PT32_LEVEL_BITS 10
109 #define PT32_LEVEL_SHIFT(level) \
110 (PAGE_SHIFT + (level - 1) * PT32_LEVEL_BITS)
112 #define PT32_LEVEL_MASK(level) \
113 (((1ULL << PT32_LEVEL_BITS) - 1) << PT32_LEVEL_SHIFT(level))
114 #define PT32_LVL_OFFSET_MASK(level) \
115 (PT32_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
116 * PT32_LEVEL_BITS))) - 1))
118 #define PT32_INDEX(address, level)\
119 (((address) >> PT32_LEVEL_SHIFT(level)) & ((1 << PT32_LEVEL_BITS) - 1))
122 #define PT64_BASE_ADDR_MASK (((1ULL << 52) - 1) & ~(u64)(PAGE_SIZE-1))
123 #define PT64_DIR_BASE_ADDR_MASK \
124 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + PT64_LEVEL_BITS)) - 1))
125 #define PT64_LVL_ADDR_MASK(level) \
126 (PT64_BASE_ADDR_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
127 * PT64_LEVEL_BITS))) - 1))
128 #define PT64_LVL_OFFSET_MASK(level) \
129 (PT64_BASE_ADDR_MASK & ((1ULL << (PAGE_SHIFT + (((level) - 1) \
130 * PT64_LEVEL_BITS))) - 1))
132 #define PT32_BASE_ADDR_MASK PAGE_MASK
133 #define PT32_DIR_BASE_ADDR_MASK \
134 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + PT32_LEVEL_BITS)) - 1))
135 #define PT32_LVL_ADDR_MASK(level) \
136 (PAGE_MASK & ~((1ULL << (PAGE_SHIFT + (((level) - 1) \
137 * PT32_LEVEL_BITS))) - 1))
139 #define PT64_PERM_MASK (PT_PRESENT_MASK | PT_WRITABLE_MASK | PT_USER_MASK \
144 #define ACC_EXEC_MASK 1
145 #define ACC_WRITE_MASK PT_WRITABLE_MASK
146 #define ACC_USER_MASK PT_USER_MASK
147 #define ACC_ALL (ACC_EXEC_MASK | ACC_WRITE_MASK | ACC_USER_MASK)
149 #include <trace/events/kvm.h>
151 #undef TRACE_INCLUDE_FILE
152 #define CREATE_TRACE_POINTS
153 #include "mmutrace.h"
155 #define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
157 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
159 struct kvm_rmap_desc
{
160 u64
*sptes
[RMAP_EXT
];
161 struct kvm_rmap_desc
*more
;
164 struct kvm_shadow_walk_iterator
{
172 #define for_each_shadow_entry(_vcpu, _addr, _walker) \
173 for (shadow_walk_init(&(_walker), _vcpu, _addr); \
174 shadow_walk_okay(&(_walker)); \
175 shadow_walk_next(&(_walker)))
178 struct kvm_unsync_walk
{
179 int (*entry
) (struct kvm_mmu_page
*sp
, struct kvm_unsync_walk
*walk
);
182 typedef int (*mmu_parent_walk_fn
) (struct kvm_vcpu
*vcpu
, struct kvm_mmu_page
*sp
);
184 static struct kmem_cache
*pte_chain_cache
;
185 static struct kmem_cache
*rmap_desc_cache
;
186 static struct kmem_cache
*mmu_page_header_cache
;
188 static u64 __read_mostly shadow_trap_nonpresent_pte
;
189 static u64 __read_mostly shadow_notrap_nonpresent_pte
;
190 static u64 __read_mostly shadow_base_present_pte
;
191 static u64 __read_mostly shadow_nx_mask
;
192 static u64 __read_mostly shadow_x_mask
; /* mutual exclusive with nx_mask */
193 static u64 __read_mostly shadow_user_mask
;
194 static u64 __read_mostly shadow_accessed_mask
;
195 static u64 __read_mostly shadow_dirty_mask
;
197 static inline u64
rsvd_bits(int s
, int e
)
199 return ((1ULL << (e
- s
+ 1)) - 1) << s
;
202 void kvm_mmu_set_nonpresent_ptes(u64 trap_pte
, u64 notrap_pte
)
204 shadow_trap_nonpresent_pte
= trap_pte
;
205 shadow_notrap_nonpresent_pte
= notrap_pte
;
207 EXPORT_SYMBOL_GPL(kvm_mmu_set_nonpresent_ptes
);
209 void kvm_mmu_set_base_ptes(u64 base_pte
)
211 shadow_base_present_pte
= base_pte
;
213 EXPORT_SYMBOL_GPL(kvm_mmu_set_base_ptes
);
215 void kvm_mmu_set_mask_ptes(u64 user_mask
, u64 accessed_mask
,
216 u64 dirty_mask
, u64 nx_mask
, u64 x_mask
)
218 shadow_user_mask
= user_mask
;
219 shadow_accessed_mask
= accessed_mask
;
220 shadow_dirty_mask
= dirty_mask
;
221 shadow_nx_mask
= nx_mask
;
222 shadow_x_mask
= x_mask
;
224 EXPORT_SYMBOL_GPL(kvm_mmu_set_mask_ptes
);
226 static int is_write_protection(struct kvm_vcpu
*vcpu
)
228 return kvm_read_cr0_bits(vcpu
, X86_CR0_WP
);
231 static int is_cpuid_PSE36(void)
236 static int is_nx(struct kvm_vcpu
*vcpu
)
238 return vcpu
->arch
.efer
& EFER_NX
;
241 static int is_shadow_present_pte(u64 pte
)
243 return pte
!= shadow_trap_nonpresent_pte
244 && pte
!= shadow_notrap_nonpresent_pte
;
247 static int is_large_pte(u64 pte
)
249 return pte
& PT_PAGE_SIZE_MASK
;
252 static int is_writable_pte(unsigned long pte
)
254 return pte
& PT_WRITABLE_MASK
;
257 static int is_dirty_gpte(unsigned long pte
)
259 return pte
& PT_DIRTY_MASK
;
262 static int is_rmap_spte(u64 pte
)
264 return is_shadow_present_pte(pte
);
267 static int is_last_spte(u64 pte
, int level
)
269 if (level
== PT_PAGE_TABLE_LEVEL
)
271 if (is_large_pte(pte
))
276 static pfn_t
spte_to_pfn(u64 pte
)
278 return (pte
& PT64_BASE_ADDR_MASK
) >> PAGE_SHIFT
;
281 static gfn_t
pse36_gfn_delta(u32 gpte
)
283 int shift
= 32 - PT32_DIR_PSE36_SHIFT
- PAGE_SHIFT
;
285 return (gpte
& PT32_DIR_PSE36_MASK
) << shift
;
288 static void __set_spte(u64
*sptep
, u64 spte
)
291 set_64bit((unsigned long *)sptep
, spte
);
293 set_64bit((unsigned long long *)sptep
, spte
);
297 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache
*cache
,
298 struct kmem_cache
*base_cache
, int min
)
302 if (cache
->nobjs
>= min
)
304 while (cache
->nobjs
< ARRAY_SIZE(cache
->objects
)) {
305 obj
= kmem_cache_zalloc(base_cache
, GFP_KERNEL
);
308 cache
->objects
[cache
->nobjs
++] = obj
;
313 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache
*mc
)
316 kfree(mc
->objects
[--mc
->nobjs
]);
319 static int mmu_topup_memory_cache_page(struct kvm_mmu_memory_cache
*cache
,
324 if (cache
->nobjs
>= min
)
326 while (cache
->nobjs
< ARRAY_SIZE(cache
->objects
)) {
327 page
= alloc_page(GFP_KERNEL
);
330 set_page_private(page
, 0);
331 cache
->objects
[cache
->nobjs
++] = page_address(page
);
336 static void mmu_free_memory_cache_page(struct kvm_mmu_memory_cache
*mc
)
339 free_page((unsigned long)mc
->objects
[--mc
->nobjs
]);
342 static int mmu_topup_memory_caches(struct kvm_vcpu
*vcpu
)
346 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_pte_chain_cache
,
350 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_rmap_desc_cache
,
354 r
= mmu_topup_memory_cache_page(&vcpu
->arch
.mmu_page_cache
, 8);
357 r
= mmu_topup_memory_cache(&vcpu
->arch
.mmu_page_header_cache
,
358 mmu_page_header_cache
, 4);
363 static void mmu_free_memory_caches(struct kvm_vcpu
*vcpu
)
365 mmu_free_memory_cache(&vcpu
->arch
.mmu_pte_chain_cache
);
366 mmu_free_memory_cache(&vcpu
->arch
.mmu_rmap_desc_cache
);
367 mmu_free_memory_cache_page(&vcpu
->arch
.mmu_page_cache
);
368 mmu_free_memory_cache(&vcpu
->arch
.mmu_page_header_cache
);
371 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache
*mc
,
377 p
= mc
->objects
[--mc
->nobjs
];
381 static struct kvm_pte_chain
*mmu_alloc_pte_chain(struct kvm_vcpu
*vcpu
)
383 return mmu_memory_cache_alloc(&vcpu
->arch
.mmu_pte_chain_cache
,
384 sizeof(struct kvm_pte_chain
));
387 static void mmu_free_pte_chain(struct kvm_pte_chain
*pc
)
392 static struct kvm_rmap_desc
*mmu_alloc_rmap_desc(struct kvm_vcpu
*vcpu
)
394 return mmu_memory_cache_alloc(&vcpu
->arch
.mmu_rmap_desc_cache
,
395 sizeof(struct kvm_rmap_desc
));
398 static void mmu_free_rmap_desc(struct kvm_rmap_desc
*rd
)
404 * Return the pointer to the largepage write count for a given
405 * gfn, handling slots that are not large page aligned.
407 static int *slot_largepage_idx(gfn_t gfn
,
408 struct kvm_memory_slot
*slot
,
413 idx
= (gfn
/ KVM_PAGES_PER_HPAGE(level
)) -
414 (slot
->base_gfn
/ KVM_PAGES_PER_HPAGE(level
));
415 return &slot
->lpage_info
[level
- 2][idx
].write_count
;
418 static void account_shadowed(struct kvm
*kvm
, gfn_t gfn
)
420 struct kvm_memory_slot
*slot
;
424 gfn
= unalias_gfn(kvm
, gfn
);
426 slot
= gfn_to_memslot_unaliased(kvm
, gfn
);
427 for (i
= PT_DIRECTORY_LEVEL
;
428 i
< PT_PAGE_TABLE_LEVEL
+ KVM_NR_PAGE_SIZES
; ++i
) {
429 write_count
= slot_largepage_idx(gfn
, slot
, i
);
434 static void unaccount_shadowed(struct kvm
*kvm
, gfn_t gfn
)
436 struct kvm_memory_slot
*slot
;
440 gfn
= unalias_gfn(kvm
, gfn
);
441 for (i
= PT_DIRECTORY_LEVEL
;
442 i
< PT_PAGE_TABLE_LEVEL
+ KVM_NR_PAGE_SIZES
; ++i
) {
443 slot
= gfn_to_memslot_unaliased(kvm
, gfn
);
444 write_count
= slot_largepage_idx(gfn
, slot
, i
);
446 WARN_ON(*write_count
< 0);
450 static int has_wrprotected_page(struct kvm
*kvm
,
454 struct kvm_memory_slot
*slot
;
457 gfn
= unalias_gfn(kvm
, gfn
);
458 slot
= gfn_to_memslot_unaliased(kvm
, gfn
);
460 largepage_idx
= slot_largepage_idx(gfn
, slot
, level
);
461 return *largepage_idx
;
467 static int host_mapping_level(struct kvm
*kvm
, gfn_t gfn
)
469 unsigned long page_size
;
472 page_size
= kvm_host_page_size(kvm
, gfn
);
474 for (i
= PT_PAGE_TABLE_LEVEL
;
475 i
< (PT_PAGE_TABLE_LEVEL
+ KVM_NR_PAGE_SIZES
); ++i
) {
476 if (page_size
>= KVM_HPAGE_SIZE(i
))
485 static int mapping_level(struct kvm_vcpu
*vcpu
, gfn_t large_gfn
)
487 struct kvm_memory_slot
*slot
;
488 int host_level
, level
, max_level
;
490 slot
= gfn_to_memslot(vcpu
->kvm
, large_gfn
);
491 if (slot
&& slot
->dirty_bitmap
)
492 return PT_PAGE_TABLE_LEVEL
;
494 host_level
= host_mapping_level(vcpu
->kvm
, large_gfn
);
496 if (host_level
== PT_PAGE_TABLE_LEVEL
)
499 max_level
= kvm_x86_ops
->get_lpage_level() < host_level
?
500 kvm_x86_ops
->get_lpage_level() : host_level
;
502 for (level
= PT_DIRECTORY_LEVEL
; level
<= max_level
; ++level
)
503 if (has_wrprotected_page(vcpu
->kvm
, large_gfn
, level
))
510 * Take gfn and return the reverse mapping to it.
511 * Note: gfn must be unaliased before this function get called
514 static unsigned long *gfn_to_rmap(struct kvm
*kvm
, gfn_t gfn
, int level
)
516 struct kvm_memory_slot
*slot
;
519 slot
= gfn_to_memslot(kvm
, gfn
);
520 if (likely(level
== PT_PAGE_TABLE_LEVEL
))
521 return &slot
->rmap
[gfn
- slot
->base_gfn
];
523 idx
= (gfn
/ KVM_PAGES_PER_HPAGE(level
)) -
524 (slot
->base_gfn
/ KVM_PAGES_PER_HPAGE(level
));
526 return &slot
->lpage_info
[level
- 2][idx
].rmap_pde
;
530 * Reverse mapping data structures:
532 * If rmapp bit zero is zero, then rmapp point to the shadw page table entry
533 * that points to page_address(page).
535 * If rmapp bit zero is one, (then rmap & ~1) points to a struct kvm_rmap_desc
536 * containing more mappings.
538 * Returns the number of rmap entries before the spte was added or zero if
539 * the spte was not added.
542 static int rmap_add(struct kvm_vcpu
*vcpu
, u64
*spte
, gfn_t gfn
)
544 struct kvm_mmu_page
*sp
;
545 struct kvm_rmap_desc
*desc
;
546 unsigned long *rmapp
;
549 if (!is_rmap_spte(*spte
))
551 gfn
= unalias_gfn(vcpu
->kvm
, gfn
);
552 sp
= page_header(__pa(spte
));
553 sp
->gfns
[spte
- sp
->spt
] = gfn
;
554 rmapp
= gfn_to_rmap(vcpu
->kvm
, gfn
, sp
->role
.level
);
556 rmap_printk("rmap_add: %p %llx 0->1\n", spte
, *spte
);
557 *rmapp
= (unsigned long)spte
;
558 } else if (!(*rmapp
& 1)) {
559 rmap_printk("rmap_add: %p %llx 1->many\n", spte
, *spte
);
560 desc
= mmu_alloc_rmap_desc(vcpu
);
561 desc
->sptes
[0] = (u64
*)*rmapp
;
562 desc
->sptes
[1] = spte
;
563 *rmapp
= (unsigned long)desc
| 1;
565 rmap_printk("rmap_add: %p %llx many->many\n", spte
, *spte
);
566 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
567 while (desc
->sptes
[RMAP_EXT
-1] && desc
->more
) {
571 if (desc
->sptes
[RMAP_EXT
-1]) {
572 desc
->more
= mmu_alloc_rmap_desc(vcpu
);
575 for (i
= 0; desc
->sptes
[i
]; ++i
)
577 desc
->sptes
[i
] = spte
;
582 static void rmap_desc_remove_entry(unsigned long *rmapp
,
583 struct kvm_rmap_desc
*desc
,
585 struct kvm_rmap_desc
*prev_desc
)
589 for (j
= RMAP_EXT
- 1; !desc
->sptes
[j
] && j
> i
; --j
)
591 desc
->sptes
[i
] = desc
->sptes
[j
];
592 desc
->sptes
[j
] = NULL
;
595 if (!prev_desc
&& !desc
->more
)
596 *rmapp
= (unsigned long)desc
->sptes
[0];
599 prev_desc
->more
= desc
->more
;
601 *rmapp
= (unsigned long)desc
->more
| 1;
602 mmu_free_rmap_desc(desc
);
605 static void rmap_remove(struct kvm
*kvm
, u64
*spte
)
607 struct kvm_rmap_desc
*desc
;
608 struct kvm_rmap_desc
*prev_desc
;
609 struct kvm_mmu_page
*sp
;
611 unsigned long *rmapp
;
614 if (!is_rmap_spte(*spte
))
616 sp
= page_header(__pa(spte
));
617 pfn
= spte_to_pfn(*spte
);
618 if (*spte
& shadow_accessed_mask
)
619 kvm_set_pfn_accessed(pfn
);
620 if (is_writable_pte(*spte
))
621 kvm_set_pfn_dirty(pfn
);
622 rmapp
= gfn_to_rmap(kvm
, sp
->gfns
[spte
- sp
->spt
], sp
->role
.level
);
624 printk(KERN_ERR
"rmap_remove: %p %llx 0->BUG\n", spte
, *spte
);
626 } else if (!(*rmapp
& 1)) {
627 rmap_printk("rmap_remove: %p %llx 1->0\n", spte
, *spte
);
628 if ((u64
*)*rmapp
!= spte
) {
629 printk(KERN_ERR
"rmap_remove: %p %llx 1->BUG\n",
635 rmap_printk("rmap_remove: %p %llx many->many\n", spte
, *spte
);
636 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
639 for (i
= 0; i
< RMAP_EXT
&& desc
->sptes
[i
]; ++i
)
640 if (desc
->sptes
[i
] == spte
) {
641 rmap_desc_remove_entry(rmapp
,
649 pr_err("rmap_remove: %p %llx many->many\n", spte
, *spte
);
654 static u64
*rmap_next(struct kvm
*kvm
, unsigned long *rmapp
, u64
*spte
)
656 struct kvm_rmap_desc
*desc
;
657 struct kvm_rmap_desc
*prev_desc
;
663 else if (!(*rmapp
& 1)) {
665 return (u64
*)*rmapp
;
668 desc
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
672 for (i
= 0; i
< RMAP_EXT
&& desc
->sptes
[i
]; ++i
) {
673 if (prev_spte
== spte
)
674 return desc
->sptes
[i
];
675 prev_spte
= desc
->sptes
[i
];
682 static int rmap_write_protect(struct kvm
*kvm
, u64 gfn
)
684 unsigned long *rmapp
;
686 int i
, write_protected
= 0;
688 gfn
= unalias_gfn(kvm
, gfn
);
689 rmapp
= gfn_to_rmap(kvm
, gfn
, PT_PAGE_TABLE_LEVEL
);
691 spte
= rmap_next(kvm
, rmapp
, NULL
);
694 BUG_ON(!(*spte
& PT_PRESENT_MASK
));
695 rmap_printk("rmap_write_protect: spte %p %llx\n", spte
, *spte
);
696 if (is_writable_pte(*spte
)) {
697 __set_spte(spte
, *spte
& ~PT_WRITABLE_MASK
);
700 spte
= rmap_next(kvm
, rmapp
, spte
);
702 if (write_protected
) {
705 spte
= rmap_next(kvm
, rmapp
, NULL
);
706 pfn
= spte_to_pfn(*spte
);
707 kvm_set_pfn_dirty(pfn
);
710 /* check for huge page mappings */
711 for (i
= PT_DIRECTORY_LEVEL
;
712 i
< PT_PAGE_TABLE_LEVEL
+ KVM_NR_PAGE_SIZES
; ++i
) {
713 rmapp
= gfn_to_rmap(kvm
, gfn
, i
);
714 spte
= rmap_next(kvm
, rmapp
, NULL
);
717 BUG_ON(!(*spte
& PT_PRESENT_MASK
));
718 BUG_ON((*spte
& (PT_PAGE_SIZE_MASK
|PT_PRESENT_MASK
)) != (PT_PAGE_SIZE_MASK
|PT_PRESENT_MASK
));
719 pgprintk("rmap_write_protect(large): spte %p %llx %lld\n", spte
, *spte
, gfn
);
720 if (is_writable_pte(*spte
)) {
721 rmap_remove(kvm
, spte
);
723 __set_spte(spte
, shadow_trap_nonpresent_pte
);
727 spte
= rmap_next(kvm
, rmapp
, spte
);
731 return write_protected
;
734 static int kvm_unmap_rmapp(struct kvm
*kvm
, unsigned long *rmapp
,
738 int need_tlb_flush
= 0;
740 while ((spte
= rmap_next(kvm
, rmapp
, NULL
))) {
741 BUG_ON(!(*spte
& PT_PRESENT_MASK
));
742 rmap_printk("kvm_rmap_unmap_hva: spte %p %llx\n", spte
, *spte
);
743 rmap_remove(kvm
, spte
);
744 __set_spte(spte
, shadow_trap_nonpresent_pte
);
747 return need_tlb_flush
;
750 static int kvm_set_pte_rmapp(struct kvm
*kvm
, unsigned long *rmapp
,
755 pte_t
*ptep
= (pte_t
*)data
;
758 WARN_ON(pte_huge(*ptep
));
759 new_pfn
= pte_pfn(*ptep
);
760 spte
= rmap_next(kvm
, rmapp
, NULL
);
762 BUG_ON(!is_shadow_present_pte(*spte
));
763 rmap_printk("kvm_set_pte_rmapp: spte %p %llx\n", spte
, *spte
);
765 if (pte_write(*ptep
)) {
766 rmap_remove(kvm
, spte
);
767 __set_spte(spte
, shadow_trap_nonpresent_pte
);
768 spte
= rmap_next(kvm
, rmapp
, NULL
);
770 new_spte
= *spte
&~ (PT64_BASE_ADDR_MASK
);
771 new_spte
|= (u64
)new_pfn
<< PAGE_SHIFT
;
773 new_spte
&= ~PT_WRITABLE_MASK
;
774 new_spte
&= ~SPTE_HOST_WRITEABLE
;
775 if (is_writable_pte(*spte
))
776 kvm_set_pfn_dirty(spte_to_pfn(*spte
));
777 __set_spte(spte
, new_spte
);
778 spte
= rmap_next(kvm
, rmapp
, spte
);
782 kvm_flush_remote_tlbs(kvm
);
787 static int kvm_handle_hva(struct kvm
*kvm
, unsigned long hva
,
789 int (*handler
)(struct kvm
*kvm
, unsigned long *rmapp
,
795 struct kvm_memslots
*slots
;
797 slots
= rcu_dereference(kvm
->memslots
);
799 for (i
= 0; i
< slots
->nmemslots
; i
++) {
800 struct kvm_memory_slot
*memslot
= &slots
->memslots
[i
];
801 unsigned long start
= memslot
->userspace_addr
;
804 end
= start
+ (memslot
->npages
<< PAGE_SHIFT
);
805 if (hva
>= start
&& hva
< end
) {
806 gfn_t gfn_offset
= (hva
- start
) >> PAGE_SHIFT
;
808 ret
= handler(kvm
, &memslot
->rmap
[gfn_offset
], data
);
810 for (j
= 0; j
< KVM_NR_PAGE_SIZES
- 1; ++j
) {
811 int idx
= gfn_offset
;
812 idx
/= KVM_PAGES_PER_HPAGE(PT_DIRECTORY_LEVEL
+ j
);
814 &memslot
->lpage_info
[j
][idx
].rmap_pde
,
817 trace_kvm_age_page(hva
, memslot
, ret
);
825 int kvm_unmap_hva(struct kvm
*kvm
, unsigned long hva
)
827 return kvm_handle_hva(kvm
, hva
, 0, kvm_unmap_rmapp
);
830 void kvm_set_spte_hva(struct kvm
*kvm
, unsigned long hva
, pte_t pte
)
832 kvm_handle_hva(kvm
, hva
, (unsigned long)&pte
, kvm_set_pte_rmapp
);
835 static int kvm_age_rmapp(struct kvm
*kvm
, unsigned long *rmapp
,
842 * Emulate the accessed bit for EPT, by checking if this page has
843 * an EPT mapping, and clearing it if it does. On the next access,
844 * a new EPT mapping will be established.
845 * This has some overhead, but not as much as the cost of swapping
846 * out actively used pages or breaking up actively used hugepages.
848 if (!shadow_accessed_mask
)
849 return kvm_unmap_rmapp(kvm
, rmapp
, data
);
851 spte
= rmap_next(kvm
, rmapp
, NULL
);
855 BUG_ON(!(_spte
& PT_PRESENT_MASK
));
856 _young
= _spte
& PT_ACCESSED_MASK
;
859 clear_bit(PT_ACCESSED_SHIFT
, (unsigned long *)spte
);
861 spte
= rmap_next(kvm
, rmapp
, spte
);
866 #define RMAP_RECYCLE_THRESHOLD 1000
868 static void rmap_recycle(struct kvm_vcpu
*vcpu
, u64
*spte
, gfn_t gfn
)
870 unsigned long *rmapp
;
871 struct kvm_mmu_page
*sp
;
873 sp
= page_header(__pa(spte
));
875 gfn
= unalias_gfn(vcpu
->kvm
, gfn
);
876 rmapp
= gfn_to_rmap(vcpu
->kvm
, gfn
, sp
->role
.level
);
878 kvm_unmap_rmapp(vcpu
->kvm
, rmapp
, 0);
879 kvm_flush_remote_tlbs(vcpu
->kvm
);
882 int kvm_age_hva(struct kvm
*kvm
, unsigned long hva
)
884 return kvm_handle_hva(kvm
, hva
, 0, kvm_age_rmapp
);
888 static int is_empty_shadow_page(u64
*spt
)
893 for (pos
= spt
, end
= pos
+ PAGE_SIZE
/ sizeof(u64
); pos
!= end
; pos
++)
894 if (is_shadow_present_pte(*pos
)) {
895 printk(KERN_ERR
"%s: %p %llx\n", __func__
,
903 static void kvm_mmu_free_page(struct kvm
*kvm
, struct kvm_mmu_page
*sp
)
905 ASSERT(is_empty_shadow_page(sp
->spt
));
907 __free_page(virt_to_page(sp
->spt
));
908 __free_page(virt_to_page(sp
->gfns
));
910 ++kvm
->arch
.n_free_mmu_pages
;
913 static unsigned kvm_page_table_hashfn(gfn_t gfn
)
915 return gfn
& ((1 << KVM_MMU_HASH_SHIFT
) - 1);
918 static struct kvm_mmu_page
*kvm_mmu_alloc_page(struct kvm_vcpu
*vcpu
,
921 struct kvm_mmu_page
*sp
;
923 sp
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_header_cache
, sizeof *sp
);
924 sp
->spt
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_cache
, PAGE_SIZE
);
925 sp
->gfns
= mmu_memory_cache_alloc(&vcpu
->arch
.mmu_page_cache
, PAGE_SIZE
);
926 set_page_private(virt_to_page(sp
->spt
), (unsigned long)sp
);
927 list_add(&sp
->link
, &vcpu
->kvm
->arch
.active_mmu_pages
);
928 INIT_LIST_HEAD(&sp
->oos_link
);
929 bitmap_zero(sp
->slot_bitmap
, KVM_MEMORY_SLOTS
+ KVM_PRIVATE_MEM_SLOTS
);
931 sp
->parent_pte
= parent_pte
;
932 --vcpu
->kvm
->arch
.n_free_mmu_pages
;
936 static void mmu_page_add_parent_pte(struct kvm_vcpu
*vcpu
,
937 struct kvm_mmu_page
*sp
, u64
*parent_pte
)
939 struct kvm_pte_chain
*pte_chain
;
940 struct hlist_node
*node
;
945 if (!sp
->multimapped
) {
946 u64
*old
= sp
->parent_pte
;
949 sp
->parent_pte
= parent_pte
;
953 pte_chain
= mmu_alloc_pte_chain(vcpu
);
954 INIT_HLIST_HEAD(&sp
->parent_ptes
);
955 hlist_add_head(&pte_chain
->link
, &sp
->parent_ptes
);
956 pte_chain
->parent_ptes
[0] = old
;
958 hlist_for_each_entry(pte_chain
, node
, &sp
->parent_ptes
, link
) {
959 if (pte_chain
->parent_ptes
[NR_PTE_CHAIN_ENTRIES
-1])
961 for (i
= 0; i
< NR_PTE_CHAIN_ENTRIES
; ++i
)
962 if (!pte_chain
->parent_ptes
[i
]) {
963 pte_chain
->parent_ptes
[i
] = parent_pte
;
967 pte_chain
= mmu_alloc_pte_chain(vcpu
);
969 hlist_add_head(&pte_chain
->link
, &sp
->parent_ptes
);
970 pte_chain
->parent_ptes
[0] = parent_pte
;
973 static void mmu_page_remove_parent_pte(struct kvm_mmu_page
*sp
,
976 struct kvm_pte_chain
*pte_chain
;
977 struct hlist_node
*node
;
980 if (!sp
->multimapped
) {
981 BUG_ON(sp
->parent_pte
!= parent_pte
);
982 sp
->parent_pte
= NULL
;
985 hlist_for_each_entry(pte_chain
, node
, &sp
->parent_ptes
, link
)
986 for (i
= 0; i
< NR_PTE_CHAIN_ENTRIES
; ++i
) {
987 if (!pte_chain
->parent_ptes
[i
])
989 if (pte_chain
->parent_ptes
[i
] != parent_pte
)
991 while (i
+ 1 < NR_PTE_CHAIN_ENTRIES
992 && pte_chain
->parent_ptes
[i
+ 1]) {
993 pte_chain
->parent_ptes
[i
]
994 = pte_chain
->parent_ptes
[i
+ 1];
997 pte_chain
->parent_ptes
[i
] = NULL
;
999 hlist_del(&pte_chain
->link
);
1000 mmu_free_pte_chain(pte_chain
);
1001 if (hlist_empty(&sp
->parent_ptes
)) {
1002 sp
->multimapped
= 0;
1003 sp
->parent_pte
= NULL
;
1012 static void mmu_parent_walk(struct kvm_vcpu
*vcpu
, struct kvm_mmu_page
*sp
,
1013 mmu_parent_walk_fn fn
)
1015 struct kvm_pte_chain
*pte_chain
;
1016 struct hlist_node
*node
;
1017 struct kvm_mmu_page
*parent_sp
;
1020 if (!sp
->multimapped
&& sp
->parent_pte
) {
1021 parent_sp
= page_header(__pa(sp
->parent_pte
));
1022 fn(vcpu
, parent_sp
);
1023 mmu_parent_walk(vcpu
, parent_sp
, fn
);
1026 hlist_for_each_entry(pte_chain
, node
, &sp
->parent_ptes
, link
)
1027 for (i
= 0; i
< NR_PTE_CHAIN_ENTRIES
; ++i
) {
1028 if (!pte_chain
->parent_ptes
[i
])
1030 parent_sp
= page_header(__pa(pte_chain
->parent_ptes
[i
]));
1031 fn(vcpu
, parent_sp
);
1032 mmu_parent_walk(vcpu
, parent_sp
, fn
);
1036 static void kvm_mmu_update_unsync_bitmap(u64
*spte
)
1039 struct kvm_mmu_page
*sp
= page_header(__pa(spte
));
1041 index
= spte
- sp
->spt
;
1042 if (!__test_and_set_bit(index
, sp
->unsync_child_bitmap
))
1043 sp
->unsync_children
++;
1044 WARN_ON(!sp
->unsync_children
);
1047 static void kvm_mmu_update_parents_unsync(struct kvm_mmu_page
*sp
)
1049 struct kvm_pte_chain
*pte_chain
;
1050 struct hlist_node
*node
;
1053 if (!sp
->parent_pte
)
1056 if (!sp
->multimapped
) {
1057 kvm_mmu_update_unsync_bitmap(sp
->parent_pte
);
1061 hlist_for_each_entry(pte_chain
, node
, &sp
->parent_ptes
, link
)
1062 for (i
= 0; i
< NR_PTE_CHAIN_ENTRIES
; ++i
) {
1063 if (!pte_chain
->parent_ptes
[i
])
1065 kvm_mmu_update_unsync_bitmap(pte_chain
->parent_ptes
[i
]);
1069 static int unsync_walk_fn(struct kvm_vcpu
*vcpu
, struct kvm_mmu_page
*sp
)
1071 kvm_mmu_update_parents_unsync(sp
);
1075 static void kvm_mmu_mark_parents_unsync(struct kvm_vcpu
*vcpu
,
1076 struct kvm_mmu_page
*sp
)
1078 mmu_parent_walk(vcpu
, sp
, unsync_walk_fn
);
1079 kvm_mmu_update_parents_unsync(sp
);
1082 static void nonpaging_prefetch_page(struct kvm_vcpu
*vcpu
,
1083 struct kvm_mmu_page
*sp
)
1087 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
)
1088 sp
->spt
[i
] = shadow_trap_nonpresent_pte
;
1091 static int nonpaging_sync_page(struct kvm_vcpu
*vcpu
,
1092 struct kvm_mmu_page
*sp
)
1097 static void nonpaging_invlpg(struct kvm_vcpu
*vcpu
, gva_t gva
)
1101 #define KVM_PAGE_ARRAY_NR 16
1103 struct kvm_mmu_pages
{
1104 struct mmu_page_and_offset
{
1105 struct kvm_mmu_page
*sp
;
1107 } page
[KVM_PAGE_ARRAY_NR
];
1111 #define for_each_unsync_children(bitmap, idx) \
1112 for (idx = find_first_bit(bitmap, 512); \
1114 idx = find_next_bit(bitmap, 512, idx+1))
1116 static int mmu_pages_add(struct kvm_mmu_pages
*pvec
, struct kvm_mmu_page
*sp
,
1122 for (i
=0; i
< pvec
->nr
; i
++)
1123 if (pvec
->page
[i
].sp
== sp
)
1126 pvec
->page
[pvec
->nr
].sp
= sp
;
1127 pvec
->page
[pvec
->nr
].idx
= idx
;
1129 return (pvec
->nr
== KVM_PAGE_ARRAY_NR
);
1132 static int __mmu_unsync_walk(struct kvm_mmu_page
*sp
,
1133 struct kvm_mmu_pages
*pvec
)
1135 int i
, ret
, nr_unsync_leaf
= 0;
1137 for_each_unsync_children(sp
->unsync_child_bitmap
, i
) {
1138 u64 ent
= sp
->spt
[i
];
1140 if (is_shadow_present_pte(ent
) && !is_large_pte(ent
)) {
1141 struct kvm_mmu_page
*child
;
1142 child
= page_header(ent
& PT64_BASE_ADDR_MASK
);
1144 if (child
->unsync_children
) {
1145 if (mmu_pages_add(pvec
, child
, i
))
1148 ret
= __mmu_unsync_walk(child
, pvec
);
1150 __clear_bit(i
, sp
->unsync_child_bitmap
);
1152 nr_unsync_leaf
+= ret
;
1157 if (child
->unsync
) {
1159 if (mmu_pages_add(pvec
, child
, i
))
1165 if (find_first_bit(sp
->unsync_child_bitmap
, 512) == 512)
1166 sp
->unsync_children
= 0;
1168 return nr_unsync_leaf
;
1171 static int mmu_unsync_walk(struct kvm_mmu_page
*sp
,
1172 struct kvm_mmu_pages
*pvec
)
1174 if (!sp
->unsync_children
)
1177 mmu_pages_add(pvec
, sp
, 0);
1178 return __mmu_unsync_walk(sp
, pvec
);
1181 static struct kvm_mmu_page
*kvm_mmu_lookup_page(struct kvm
*kvm
, gfn_t gfn
)
1184 struct hlist_head
*bucket
;
1185 struct kvm_mmu_page
*sp
;
1186 struct hlist_node
*node
;
1188 pgprintk("%s: looking for gfn %lx\n", __func__
, gfn
);
1189 index
= kvm_page_table_hashfn(gfn
);
1190 bucket
= &kvm
->arch
.mmu_page_hash
[index
];
1191 hlist_for_each_entry(sp
, node
, bucket
, hash_link
)
1192 if (sp
->gfn
== gfn
&& !sp
->role
.direct
1193 && !sp
->role
.invalid
) {
1194 pgprintk("%s: found role %x\n",
1195 __func__
, sp
->role
.word
);
1201 static void kvm_unlink_unsync_page(struct kvm
*kvm
, struct kvm_mmu_page
*sp
)
1203 WARN_ON(!sp
->unsync
);
1205 --kvm
->stat
.mmu_unsync
;
1208 static int kvm_mmu_zap_page(struct kvm
*kvm
, struct kvm_mmu_page
*sp
);
1210 static int kvm_sync_page(struct kvm_vcpu
*vcpu
, struct kvm_mmu_page
*sp
)
1212 if (sp
->role
.glevels
!= vcpu
->arch
.mmu
.root_level
) {
1213 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1217 trace_kvm_mmu_sync_page(sp
);
1218 if (rmap_write_protect(vcpu
->kvm
, sp
->gfn
))
1219 kvm_flush_remote_tlbs(vcpu
->kvm
);
1220 kvm_unlink_unsync_page(vcpu
->kvm
, sp
);
1221 if (vcpu
->arch
.mmu
.sync_page(vcpu
, sp
)) {
1222 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
1226 kvm_mmu_flush_tlb(vcpu
);
1230 struct mmu_page_path
{
1231 struct kvm_mmu_page
*parent
[PT64_ROOT_LEVEL
-1];
1232 unsigned int idx
[PT64_ROOT_LEVEL
-1];
1235 #define for_each_sp(pvec, sp, parents, i) \
1236 for (i = mmu_pages_next(&pvec, &parents, -1), \
1237 sp = pvec.page[i].sp; \
1238 i < pvec.nr && ({ sp = pvec.page[i].sp; 1;}); \
1239 i = mmu_pages_next(&pvec, &parents, i))
1241 static int mmu_pages_next(struct kvm_mmu_pages
*pvec
,
1242 struct mmu_page_path
*parents
,
1247 for (n
= i
+1; n
< pvec
->nr
; n
++) {
1248 struct kvm_mmu_page
*sp
= pvec
->page
[n
].sp
;
1250 if (sp
->role
.level
== PT_PAGE_TABLE_LEVEL
) {
1251 parents
->idx
[0] = pvec
->page
[n
].idx
;
1255 parents
->parent
[sp
->role
.level
-2] = sp
;
1256 parents
->idx
[sp
->role
.level
-1] = pvec
->page
[n
].idx
;
1262 static void mmu_pages_clear_parents(struct mmu_page_path
*parents
)
1264 struct kvm_mmu_page
*sp
;
1265 unsigned int level
= 0;
1268 unsigned int idx
= parents
->idx
[level
];
1270 sp
= parents
->parent
[level
];
1274 --sp
->unsync_children
;
1275 WARN_ON((int)sp
->unsync_children
< 0);
1276 __clear_bit(idx
, sp
->unsync_child_bitmap
);
1278 } while (level
< PT64_ROOT_LEVEL
-1 && !sp
->unsync_children
);
1281 static void kvm_mmu_pages_init(struct kvm_mmu_page
*parent
,
1282 struct mmu_page_path
*parents
,
1283 struct kvm_mmu_pages
*pvec
)
1285 parents
->parent
[parent
->role
.level
-1] = NULL
;
1289 static void mmu_sync_children(struct kvm_vcpu
*vcpu
,
1290 struct kvm_mmu_page
*parent
)
1293 struct kvm_mmu_page
*sp
;
1294 struct mmu_page_path parents
;
1295 struct kvm_mmu_pages pages
;
1297 kvm_mmu_pages_init(parent
, &parents
, &pages
);
1298 while (mmu_unsync_walk(parent
, &pages
)) {
1301 for_each_sp(pages
, sp
, parents
, i
)
1302 protected |= rmap_write_protect(vcpu
->kvm
, sp
->gfn
);
1305 kvm_flush_remote_tlbs(vcpu
->kvm
);
1307 for_each_sp(pages
, sp
, parents
, i
) {
1308 kvm_sync_page(vcpu
, sp
);
1309 mmu_pages_clear_parents(&parents
);
1311 cond_resched_lock(&vcpu
->kvm
->mmu_lock
);
1312 kvm_mmu_pages_init(parent
, &parents
, &pages
);
1316 static struct kvm_mmu_page
*kvm_mmu_get_page(struct kvm_vcpu
*vcpu
,
1324 union kvm_mmu_page_role role
;
1327 struct hlist_head
*bucket
;
1328 struct kvm_mmu_page
*sp
;
1329 struct hlist_node
*node
, *tmp
;
1331 role
= vcpu
->arch
.mmu
.base_role
;
1333 role
.direct
= direct
;
1334 role
.access
= access
;
1335 if (vcpu
->arch
.mmu
.root_level
<= PT32_ROOT_LEVEL
) {
1336 quadrant
= gaddr
>> (PAGE_SHIFT
+ (PT64_PT_BITS
* level
));
1337 quadrant
&= (1 << ((PT32_PT_BITS
- PT64_PT_BITS
) * level
)) - 1;
1338 role
.quadrant
= quadrant
;
1340 index
= kvm_page_table_hashfn(gfn
);
1341 bucket
= &vcpu
->kvm
->arch
.mmu_page_hash
[index
];
1342 hlist_for_each_entry_safe(sp
, node
, tmp
, bucket
, hash_link
)
1343 if (sp
->gfn
== gfn
) {
1345 if (kvm_sync_page(vcpu
, sp
))
1348 if (sp
->role
.word
!= role
.word
)
1351 mmu_page_add_parent_pte(vcpu
, sp
, parent_pte
);
1352 if (sp
->unsync_children
) {
1353 set_bit(KVM_REQ_MMU_SYNC
, &vcpu
->requests
);
1354 kvm_mmu_mark_parents_unsync(vcpu
, sp
);
1356 trace_kvm_mmu_get_page(sp
, false);
1359 ++vcpu
->kvm
->stat
.mmu_cache_miss
;
1360 sp
= kvm_mmu_alloc_page(vcpu
, parent_pte
);
1365 hlist_add_head(&sp
->hash_link
, bucket
);
1367 if (rmap_write_protect(vcpu
->kvm
, gfn
))
1368 kvm_flush_remote_tlbs(vcpu
->kvm
);
1369 account_shadowed(vcpu
->kvm
, gfn
);
1371 if (shadow_trap_nonpresent_pte
!= shadow_notrap_nonpresent_pte
)
1372 vcpu
->arch
.mmu
.prefetch_page(vcpu
, sp
);
1374 nonpaging_prefetch_page(vcpu
, sp
);
1375 trace_kvm_mmu_get_page(sp
, true);
1379 static void shadow_walk_init(struct kvm_shadow_walk_iterator
*iterator
,
1380 struct kvm_vcpu
*vcpu
, u64 addr
)
1382 iterator
->addr
= addr
;
1383 iterator
->shadow_addr
= vcpu
->arch
.mmu
.root_hpa
;
1384 iterator
->level
= vcpu
->arch
.mmu
.shadow_root_level
;
1385 if (iterator
->level
== PT32E_ROOT_LEVEL
) {
1386 iterator
->shadow_addr
1387 = vcpu
->arch
.mmu
.pae_root
[(addr
>> 30) & 3];
1388 iterator
->shadow_addr
&= PT64_BASE_ADDR_MASK
;
1390 if (!iterator
->shadow_addr
)
1391 iterator
->level
= 0;
1395 static bool shadow_walk_okay(struct kvm_shadow_walk_iterator
*iterator
)
1397 if (iterator
->level
< PT_PAGE_TABLE_LEVEL
)
1400 if (iterator
->level
== PT_PAGE_TABLE_LEVEL
)
1401 if (is_large_pte(*iterator
->sptep
))
1404 iterator
->index
= SHADOW_PT_INDEX(iterator
->addr
, iterator
->level
);
1405 iterator
->sptep
= ((u64
*)__va(iterator
->shadow_addr
)) + iterator
->index
;
1409 static void shadow_walk_next(struct kvm_shadow_walk_iterator
*iterator
)
1411 iterator
->shadow_addr
= *iterator
->sptep
& PT64_BASE_ADDR_MASK
;
1415 static void kvm_mmu_page_unlink_children(struct kvm
*kvm
,
1416 struct kvm_mmu_page
*sp
)
1424 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
1427 if (is_shadow_present_pte(ent
)) {
1428 if (!is_last_spte(ent
, sp
->role
.level
)) {
1429 ent
&= PT64_BASE_ADDR_MASK
;
1430 mmu_page_remove_parent_pte(page_header(ent
),
1433 if (is_large_pte(ent
))
1435 rmap_remove(kvm
, &pt
[i
]);
1438 pt
[i
] = shadow_trap_nonpresent_pte
;
1442 static void kvm_mmu_put_page(struct kvm_mmu_page
*sp
, u64
*parent_pte
)
1444 mmu_page_remove_parent_pte(sp
, parent_pte
);
1447 static void kvm_mmu_reset_last_pte_updated(struct kvm
*kvm
)
1450 struct kvm_vcpu
*vcpu
;
1452 kvm_for_each_vcpu(i
, vcpu
, kvm
)
1453 vcpu
->arch
.last_pte_updated
= NULL
;
1456 static void kvm_mmu_unlink_parents(struct kvm
*kvm
, struct kvm_mmu_page
*sp
)
1460 while (sp
->multimapped
|| sp
->parent_pte
) {
1461 if (!sp
->multimapped
)
1462 parent_pte
= sp
->parent_pte
;
1464 struct kvm_pte_chain
*chain
;
1466 chain
= container_of(sp
->parent_ptes
.first
,
1467 struct kvm_pte_chain
, link
);
1468 parent_pte
= chain
->parent_ptes
[0];
1470 BUG_ON(!parent_pte
);
1471 kvm_mmu_put_page(sp
, parent_pte
);
1472 __set_spte(parent_pte
, shadow_trap_nonpresent_pte
);
1476 static int mmu_zap_unsync_children(struct kvm
*kvm
,
1477 struct kvm_mmu_page
*parent
)
1480 struct mmu_page_path parents
;
1481 struct kvm_mmu_pages pages
;
1483 if (parent
->role
.level
== PT_PAGE_TABLE_LEVEL
)
1486 kvm_mmu_pages_init(parent
, &parents
, &pages
);
1487 while (mmu_unsync_walk(parent
, &pages
)) {
1488 struct kvm_mmu_page
*sp
;
1490 for_each_sp(pages
, sp
, parents
, i
) {
1491 kvm_mmu_zap_page(kvm
, sp
);
1492 mmu_pages_clear_parents(&parents
);
1495 kvm_mmu_pages_init(parent
, &parents
, &pages
);
1501 static int kvm_mmu_zap_page(struct kvm
*kvm
, struct kvm_mmu_page
*sp
)
1505 trace_kvm_mmu_zap_page(sp
);
1506 ++kvm
->stat
.mmu_shadow_zapped
;
1507 ret
= mmu_zap_unsync_children(kvm
, sp
);
1508 kvm_mmu_page_unlink_children(kvm
, sp
);
1509 kvm_mmu_unlink_parents(kvm
, sp
);
1510 kvm_flush_remote_tlbs(kvm
);
1511 if (!sp
->role
.invalid
&& !sp
->role
.direct
)
1512 unaccount_shadowed(kvm
, sp
->gfn
);
1514 kvm_unlink_unsync_page(kvm
, sp
);
1515 if (!sp
->root_count
) {
1516 hlist_del(&sp
->hash_link
);
1517 kvm_mmu_free_page(kvm
, sp
);
1519 sp
->role
.invalid
= 1;
1520 list_move(&sp
->link
, &kvm
->arch
.active_mmu_pages
);
1521 kvm_reload_remote_mmus(kvm
);
1523 kvm_mmu_reset_last_pte_updated(kvm
);
1528 * Changing the number of mmu pages allocated to the vm
1529 * Note: if kvm_nr_mmu_pages is too small, you will get dead lock
1531 void kvm_mmu_change_mmu_pages(struct kvm
*kvm
, unsigned int kvm_nr_mmu_pages
)
1535 used_pages
= kvm
->arch
.n_alloc_mmu_pages
- kvm
->arch
.n_free_mmu_pages
;
1536 used_pages
= max(0, used_pages
);
1539 * If we set the number of mmu pages to be smaller be than the
1540 * number of actived pages , we must to free some mmu pages before we
1544 if (used_pages
> kvm_nr_mmu_pages
) {
1545 while (used_pages
> kvm_nr_mmu_pages
) {
1546 struct kvm_mmu_page
*page
;
1548 page
= container_of(kvm
->arch
.active_mmu_pages
.prev
,
1549 struct kvm_mmu_page
, link
);
1550 kvm_mmu_zap_page(kvm
, page
);
1553 kvm
->arch
.n_free_mmu_pages
= 0;
1556 kvm
->arch
.n_free_mmu_pages
+= kvm_nr_mmu_pages
1557 - kvm
->arch
.n_alloc_mmu_pages
;
1559 kvm
->arch
.n_alloc_mmu_pages
= kvm_nr_mmu_pages
;
1562 static int kvm_mmu_unprotect_page(struct kvm
*kvm
, gfn_t gfn
)
1565 struct hlist_head
*bucket
;
1566 struct kvm_mmu_page
*sp
;
1567 struct hlist_node
*node
, *n
;
1570 pgprintk("%s: looking for gfn %lx\n", __func__
, gfn
);
1572 index
= kvm_page_table_hashfn(gfn
);
1573 bucket
= &kvm
->arch
.mmu_page_hash
[index
];
1574 hlist_for_each_entry_safe(sp
, node
, n
, bucket
, hash_link
)
1575 if (sp
->gfn
== gfn
&& !sp
->role
.direct
) {
1576 pgprintk("%s: gfn %lx role %x\n", __func__
, gfn
,
1579 if (kvm_mmu_zap_page(kvm
, sp
))
1585 static void mmu_unshadow(struct kvm
*kvm
, gfn_t gfn
)
1588 struct hlist_head
*bucket
;
1589 struct kvm_mmu_page
*sp
;
1590 struct hlist_node
*node
, *nn
;
1592 index
= kvm_page_table_hashfn(gfn
);
1593 bucket
= &kvm
->arch
.mmu_page_hash
[index
];
1594 hlist_for_each_entry_safe(sp
, node
, nn
, bucket
, hash_link
) {
1595 if (sp
->gfn
== gfn
&& !sp
->role
.direct
1596 && !sp
->role
.invalid
) {
1597 pgprintk("%s: zap %lx %x\n",
1598 __func__
, gfn
, sp
->role
.word
);
1599 kvm_mmu_zap_page(kvm
, sp
);
1604 static void page_header_update_slot(struct kvm
*kvm
, void *pte
, gfn_t gfn
)
1606 int slot
= memslot_id(kvm
, gfn
);
1607 struct kvm_mmu_page
*sp
= page_header(__pa(pte
));
1609 __set_bit(slot
, sp
->slot_bitmap
);
1612 static void mmu_convert_notrap(struct kvm_mmu_page
*sp
)
1617 if (shadow_trap_nonpresent_pte
== shadow_notrap_nonpresent_pte
)
1620 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
1621 if (pt
[i
] == shadow_notrap_nonpresent_pte
)
1622 __set_spte(&pt
[i
], shadow_trap_nonpresent_pte
);
1626 struct page
*gva_to_page(struct kvm_vcpu
*vcpu
, gva_t gva
)
1630 gpa_t gpa
= kvm_mmu_gva_to_gpa_read(vcpu
, gva
, NULL
);
1632 if (gpa
== UNMAPPED_GVA
)
1635 page
= gfn_to_page(vcpu
->kvm
, gpa
>> PAGE_SHIFT
);
1641 * The function is based on mtrr_type_lookup() in
1642 * arch/x86/kernel/cpu/mtrr/generic.c
1644 static int get_mtrr_type(struct mtrr_state_type
*mtrr_state
,
1649 u8 prev_match
, curr_match
;
1650 int num_var_ranges
= KVM_NR_VAR_MTRR
;
1652 if (!mtrr_state
->enabled
)
1655 /* Make end inclusive end, instead of exclusive */
1658 /* Look in fixed ranges. Just return the type as per start */
1659 if (mtrr_state
->have_fixed
&& (start
< 0x100000)) {
1662 if (start
< 0x80000) {
1664 idx
+= (start
>> 16);
1665 return mtrr_state
->fixed_ranges
[idx
];
1666 } else if (start
< 0xC0000) {
1668 idx
+= ((start
- 0x80000) >> 14);
1669 return mtrr_state
->fixed_ranges
[idx
];
1670 } else if (start
< 0x1000000) {
1672 idx
+= ((start
- 0xC0000) >> 12);
1673 return mtrr_state
->fixed_ranges
[idx
];
1678 * Look in variable ranges
1679 * Look of multiple ranges matching this address and pick type
1680 * as per MTRR precedence
1682 if (!(mtrr_state
->enabled
& 2))
1683 return mtrr_state
->def_type
;
1686 for (i
= 0; i
< num_var_ranges
; ++i
) {
1687 unsigned short start_state
, end_state
;
1689 if (!(mtrr_state
->var_ranges
[i
].mask_lo
& (1 << 11)))
1692 base
= (((u64
)mtrr_state
->var_ranges
[i
].base_hi
) << 32) +
1693 (mtrr_state
->var_ranges
[i
].base_lo
& PAGE_MASK
);
1694 mask
= (((u64
)mtrr_state
->var_ranges
[i
].mask_hi
) << 32) +
1695 (mtrr_state
->var_ranges
[i
].mask_lo
& PAGE_MASK
);
1697 start_state
= ((start
& mask
) == (base
& mask
));
1698 end_state
= ((end
& mask
) == (base
& mask
));
1699 if (start_state
!= end_state
)
1702 if ((start
& mask
) != (base
& mask
))
1705 curr_match
= mtrr_state
->var_ranges
[i
].base_lo
& 0xff;
1706 if (prev_match
== 0xFF) {
1707 prev_match
= curr_match
;
1711 if (prev_match
== MTRR_TYPE_UNCACHABLE
||
1712 curr_match
== MTRR_TYPE_UNCACHABLE
)
1713 return MTRR_TYPE_UNCACHABLE
;
1715 if ((prev_match
== MTRR_TYPE_WRBACK
&&
1716 curr_match
== MTRR_TYPE_WRTHROUGH
) ||
1717 (prev_match
== MTRR_TYPE_WRTHROUGH
&&
1718 curr_match
== MTRR_TYPE_WRBACK
)) {
1719 prev_match
= MTRR_TYPE_WRTHROUGH
;
1720 curr_match
= MTRR_TYPE_WRTHROUGH
;
1723 if (prev_match
!= curr_match
)
1724 return MTRR_TYPE_UNCACHABLE
;
1727 if (prev_match
!= 0xFF)
1730 return mtrr_state
->def_type
;
1733 u8
kvm_get_guest_memory_type(struct kvm_vcpu
*vcpu
, gfn_t gfn
)
1737 mtrr
= get_mtrr_type(&vcpu
->arch
.mtrr_state
, gfn
<< PAGE_SHIFT
,
1738 (gfn
<< PAGE_SHIFT
) + PAGE_SIZE
);
1739 if (mtrr
== 0xfe || mtrr
== 0xff)
1740 mtrr
= MTRR_TYPE_WRBACK
;
1743 EXPORT_SYMBOL_GPL(kvm_get_guest_memory_type
);
1745 static int kvm_unsync_page(struct kvm_vcpu
*vcpu
, struct kvm_mmu_page
*sp
)
1748 struct hlist_head
*bucket
;
1749 struct kvm_mmu_page
*s
;
1750 struct hlist_node
*node
, *n
;
1752 trace_kvm_mmu_unsync_page(sp
);
1753 index
= kvm_page_table_hashfn(sp
->gfn
);
1754 bucket
= &vcpu
->kvm
->arch
.mmu_page_hash
[index
];
1755 /* don't unsync if pagetable is shadowed with multiple roles */
1756 hlist_for_each_entry_safe(s
, node
, n
, bucket
, hash_link
) {
1757 if (s
->gfn
!= sp
->gfn
|| s
->role
.direct
)
1759 if (s
->role
.word
!= sp
->role
.word
)
1762 ++vcpu
->kvm
->stat
.mmu_unsync
;
1765 kvm_mmu_mark_parents_unsync(vcpu
, sp
);
1767 mmu_convert_notrap(sp
);
1771 static int mmu_need_write_protect(struct kvm_vcpu
*vcpu
, gfn_t gfn
,
1774 struct kvm_mmu_page
*shadow
;
1776 shadow
= kvm_mmu_lookup_page(vcpu
->kvm
, gfn
);
1778 if (shadow
->role
.level
!= PT_PAGE_TABLE_LEVEL
)
1782 if (can_unsync
&& oos_shadow
)
1783 return kvm_unsync_page(vcpu
, shadow
);
1789 static int set_spte(struct kvm_vcpu
*vcpu
, u64
*sptep
,
1790 unsigned pte_access
, int user_fault
,
1791 int write_fault
, int dirty
, int level
,
1792 gfn_t gfn
, pfn_t pfn
, bool speculative
,
1793 bool can_unsync
, bool reset_host_protection
)
1799 * We don't set the accessed bit, since we sometimes want to see
1800 * whether the guest actually used the pte (in order to detect
1803 spte
= shadow_base_present_pte
| shadow_dirty_mask
;
1805 spte
|= shadow_accessed_mask
;
1807 pte_access
&= ~ACC_WRITE_MASK
;
1808 if (pte_access
& ACC_EXEC_MASK
)
1809 spte
|= shadow_x_mask
;
1811 spte
|= shadow_nx_mask
;
1812 if (pte_access
& ACC_USER_MASK
)
1813 spte
|= shadow_user_mask
;
1814 if (level
> PT_PAGE_TABLE_LEVEL
)
1815 spte
|= PT_PAGE_SIZE_MASK
;
1817 spte
|= kvm_x86_ops
->get_mt_mask(vcpu
, gfn
,
1818 kvm_is_mmio_pfn(pfn
));
1820 if (reset_host_protection
)
1821 spte
|= SPTE_HOST_WRITEABLE
;
1823 spte
|= (u64
)pfn
<< PAGE_SHIFT
;
1825 if ((pte_access
& ACC_WRITE_MASK
)
1826 || (write_fault
&& !is_write_protection(vcpu
) && !user_fault
)) {
1828 if (level
> PT_PAGE_TABLE_LEVEL
&&
1829 has_wrprotected_page(vcpu
->kvm
, gfn
, level
)) {
1831 spte
= shadow_trap_nonpresent_pte
;
1835 spte
|= PT_WRITABLE_MASK
;
1838 * Optimization: for pte sync, if spte was writable the hash
1839 * lookup is unnecessary (and expensive). Write protection
1840 * is responsibility of mmu_get_page / kvm_sync_page.
1841 * Same reasoning can be applied to dirty page accounting.
1843 if (!can_unsync
&& is_writable_pte(*sptep
))
1846 if (mmu_need_write_protect(vcpu
, gfn
, can_unsync
)) {
1847 pgprintk("%s: found shadow page for %lx, marking ro\n",
1850 pte_access
&= ~ACC_WRITE_MASK
;
1851 if (is_writable_pte(spte
))
1852 spte
&= ~PT_WRITABLE_MASK
;
1856 if (pte_access
& ACC_WRITE_MASK
)
1857 mark_page_dirty(vcpu
->kvm
, gfn
);
1860 __set_spte(sptep
, spte
);
1864 static void mmu_set_spte(struct kvm_vcpu
*vcpu
, u64
*sptep
,
1865 unsigned pt_access
, unsigned pte_access
,
1866 int user_fault
, int write_fault
, int dirty
,
1867 int *ptwrite
, int level
, gfn_t gfn
,
1868 pfn_t pfn
, bool speculative
,
1869 bool reset_host_protection
)
1871 int was_rmapped
= 0;
1872 int was_writable
= is_writable_pte(*sptep
);
1875 pgprintk("%s: spte %llx access %x write_fault %d"
1876 " user_fault %d gfn %lx\n",
1877 __func__
, *sptep
, pt_access
,
1878 write_fault
, user_fault
, gfn
);
1880 if (is_rmap_spte(*sptep
)) {
1882 * If we overwrite a PTE page pointer with a 2MB PMD, unlink
1883 * the parent of the now unreachable PTE.
1885 if (level
> PT_PAGE_TABLE_LEVEL
&&
1886 !is_large_pte(*sptep
)) {
1887 struct kvm_mmu_page
*child
;
1890 child
= page_header(pte
& PT64_BASE_ADDR_MASK
);
1891 mmu_page_remove_parent_pte(child
, sptep
);
1892 } else if (pfn
!= spte_to_pfn(*sptep
)) {
1893 pgprintk("hfn old %lx new %lx\n",
1894 spte_to_pfn(*sptep
), pfn
);
1895 rmap_remove(vcpu
->kvm
, sptep
);
1900 if (set_spte(vcpu
, sptep
, pte_access
, user_fault
, write_fault
,
1901 dirty
, level
, gfn
, pfn
, speculative
, true,
1902 reset_host_protection
)) {
1905 kvm_x86_ops
->tlb_flush(vcpu
);
1908 pgprintk("%s: setting spte %llx\n", __func__
, *sptep
);
1909 pgprintk("instantiating %s PTE (%s) at %ld (%llx) addr %p\n",
1910 is_large_pte(*sptep
)? "2MB" : "4kB",
1911 *sptep
& PT_PRESENT_MASK
?"RW":"R", gfn
,
1913 if (!was_rmapped
&& is_large_pte(*sptep
))
1914 ++vcpu
->kvm
->stat
.lpages
;
1916 page_header_update_slot(vcpu
->kvm
, sptep
, gfn
);
1918 rmap_count
= rmap_add(vcpu
, sptep
, gfn
);
1919 kvm_release_pfn_clean(pfn
);
1920 if (rmap_count
> RMAP_RECYCLE_THRESHOLD
)
1921 rmap_recycle(vcpu
, sptep
, gfn
);
1924 kvm_release_pfn_dirty(pfn
);
1926 kvm_release_pfn_clean(pfn
);
1929 vcpu
->arch
.last_pte_updated
= sptep
;
1930 vcpu
->arch
.last_pte_gfn
= gfn
;
1934 static void nonpaging_new_cr3(struct kvm_vcpu
*vcpu
)
1938 static int __direct_map(struct kvm_vcpu
*vcpu
, gpa_t v
, int write
,
1939 int level
, gfn_t gfn
, pfn_t pfn
)
1941 struct kvm_shadow_walk_iterator iterator
;
1942 struct kvm_mmu_page
*sp
;
1946 for_each_shadow_entry(vcpu
, (u64
)gfn
<< PAGE_SHIFT
, iterator
) {
1947 if (iterator
.level
== level
) {
1948 mmu_set_spte(vcpu
, iterator
.sptep
, ACC_ALL
, ACC_ALL
,
1949 0, write
, 1, &pt_write
,
1950 level
, gfn
, pfn
, false, true);
1951 ++vcpu
->stat
.pf_fixed
;
1955 if (*iterator
.sptep
== shadow_trap_nonpresent_pte
) {
1956 pseudo_gfn
= (iterator
.addr
& PT64_DIR_BASE_ADDR_MASK
) >> PAGE_SHIFT
;
1957 sp
= kvm_mmu_get_page(vcpu
, pseudo_gfn
, iterator
.addr
,
1959 1, ACC_ALL
, iterator
.sptep
);
1961 pgprintk("nonpaging_map: ENOMEM\n");
1962 kvm_release_pfn_clean(pfn
);
1966 __set_spte(iterator
.sptep
,
1968 | PT_PRESENT_MASK
| PT_WRITABLE_MASK
1969 | shadow_user_mask
| shadow_x_mask
);
1975 static int nonpaging_map(struct kvm_vcpu
*vcpu
, gva_t v
, int write
, gfn_t gfn
)
1980 unsigned long mmu_seq
;
1982 level
= mapping_level(vcpu
, gfn
);
1985 * This path builds a PAE pagetable - so we can map 2mb pages at
1986 * maximum. Therefore check if the level is larger than that.
1988 if (level
> PT_DIRECTORY_LEVEL
)
1989 level
= PT_DIRECTORY_LEVEL
;
1991 gfn
&= ~(KVM_PAGES_PER_HPAGE(level
) - 1);
1993 mmu_seq
= vcpu
->kvm
->mmu_notifier_seq
;
1995 pfn
= gfn_to_pfn(vcpu
->kvm
, gfn
);
1998 if (is_error_pfn(pfn
)) {
1999 kvm_release_pfn_clean(pfn
);
2003 spin_lock(&vcpu
->kvm
->mmu_lock
);
2004 if (mmu_notifier_retry(vcpu
, mmu_seq
))
2006 kvm_mmu_free_some_pages(vcpu
);
2007 r
= __direct_map(vcpu
, v
, write
, level
, gfn
, pfn
);
2008 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2014 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2015 kvm_release_pfn_clean(pfn
);
2020 static void mmu_free_roots(struct kvm_vcpu
*vcpu
)
2023 struct kvm_mmu_page
*sp
;
2025 if (!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
))
2027 spin_lock(&vcpu
->kvm
->mmu_lock
);
2028 if (vcpu
->arch
.mmu
.shadow_root_level
== PT64_ROOT_LEVEL
) {
2029 hpa_t root
= vcpu
->arch
.mmu
.root_hpa
;
2031 sp
= page_header(root
);
2033 if (!sp
->root_count
&& sp
->role
.invalid
)
2034 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
2035 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
2036 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2039 for (i
= 0; i
< 4; ++i
) {
2040 hpa_t root
= vcpu
->arch
.mmu
.pae_root
[i
];
2043 root
&= PT64_BASE_ADDR_MASK
;
2044 sp
= page_header(root
);
2046 if (!sp
->root_count
&& sp
->role
.invalid
)
2047 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
2049 vcpu
->arch
.mmu
.pae_root
[i
] = INVALID_PAGE
;
2051 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2052 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
2055 static int mmu_check_root(struct kvm_vcpu
*vcpu
, gfn_t root_gfn
)
2059 if (!kvm_is_visible_gfn(vcpu
->kvm
, root_gfn
)) {
2060 set_bit(KVM_REQ_TRIPLE_FAULT
, &vcpu
->requests
);
2067 static int mmu_alloc_roots(struct kvm_vcpu
*vcpu
)
2071 struct kvm_mmu_page
*sp
;
2075 root_gfn
= vcpu
->arch
.cr3
>> PAGE_SHIFT
;
2077 if (vcpu
->arch
.mmu
.shadow_root_level
== PT64_ROOT_LEVEL
) {
2078 hpa_t root
= vcpu
->arch
.mmu
.root_hpa
;
2080 ASSERT(!VALID_PAGE(root
));
2083 if (mmu_check_root(vcpu
, root_gfn
))
2085 sp
= kvm_mmu_get_page(vcpu
, root_gfn
, 0,
2086 PT64_ROOT_LEVEL
, direct
,
2088 root
= __pa(sp
->spt
);
2090 vcpu
->arch
.mmu
.root_hpa
= root
;
2093 direct
= !is_paging(vcpu
);
2096 for (i
= 0; i
< 4; ++i
) {
2097 hpa_t root
= vcpu
->arch
.mmu
.pae_root
[i
];
2099 ASSERT(!VALID_PAGE(root
));
2100 if (vcpu
->arch
.mmu
.root_level
== PT32E_ROOT_LEVEL
) {
2101 pdptr
= kvm_pdptr_read(vcpu
, i
);
2102 if (!is_present_gpte(pdptr
)) {
2103 vcpu
->arch
.mmu
.pae_root
[i
] = 0;
2106 root_gfn
= pdptr
>> PAGE_SHIFT
;
2107 } else if (vcpu
->arch
.mmu
.root_level
== 0)
2109 if (mmu_check_root(vcpu
, root_gfn
))
2111 sp
= kvm_mmu_get_page(vcpu
, root_gfn
, i
<< 30,
2112 PT32_ROOT_LEVEL
, direct
,
2114 root
= __pa(sp
->spt
);
2116 vcpu
->arch
.mmu
.pae_root
[i
] = root
| PT_PRESENT_MASK
;
2118 vcpu
->arch
.mmu
.root_hpa
= __pa(vcpu
->arch
.mmu
.pae_root
);
2122 static void mmu_sync_roots(struct kvm_vcpu
*vcpu
)
2125 struct kvm_mmu_page
*sp
;
2127 if (!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
))
2129 if (vcpu
->arch
.mmu
.shadow_root_level
== PT64_ROOT_LEVEL
) {
2130 hpa_t root
= vcpu
->arch
.mmu
.root_hpa
;
2131 sp
= page_header(root
);
2132 mmu_sync_children(vcpu
, sp
);
2135 for (i
= 0; i
< 4; ++i
) {
2136 hpa_t root
= vcpu
->arch
.mmu
.pae_root
[i
];
2138 if (root
&& VALID_PAGE(root
)) {
2139 root
&= PT64_BASE_ADDR_MASK
;
2140 sp
= page_header(root
);
2141 mmu_sync_children(vcpu
, sp
);
2146 void kvm_mmu_sync_roots(struct kvm_vcpu
*vcpu
)
2148 spin_lock(&vcpu
->kvm
->mmu_lock
);
2149 mmu_sync_roots(vcpu
);
2150 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2153 static gpa_t
nonpaging_gva_to_gpa(struct kvm_vcpu
*vcpu
, gva_t vaddr
,
2154 u32 access
, u32
*error
)
2161 static int nonpaging_page_fault(struct kvm_vcpu
*vcpu
, gva_t gva
,
2167 pgprintk("%s: gva %lx error %x\n", __func__
, gva
, error_code
);
2168 r
= mmu_topup_memory_caches(vcpu
);
2173 ASSERT(VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
2175 gfn
= gva
>> PAGE_SHIFT
;
2177 return nonpaging_map(vcpu
, gva
& PAGE_MASK
,
2178 error_code
& PFERR_WRITE_MASK
, gfn
);
2181 static int tdp_page_fault(struct kvm_vcpu
*vcpu
, gva_t gpa
,
2187 gfn_t gfn
= gpa
>> PAGE_SHIFT
;
2188 unsigned long mmu_seq
;
2191 ASSERT(VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
2193 r
= mmu_topup_memory_caches(vcpu
);
2197 level
= mapping_level(vcpu
, gfn
);
2199 gfn
&= ~(KVM_PAGES_PER_HPAGE(level
) - 1);
2201 mmu_seq
= vcpu
->kvm
->mmu_notifier_seq
;
2203 pfn
= gfn_to_pfn(vcpu
->kvm
, gfn
);
2204 if (is_error_pfn(pfn
)) {
2205 kvm_release_pfn_clean(pfn
);
2208 spin_lock(&vcpu
->kvm
->mmu_lock
);
2209 if (mmu_notifier_retry(vcpu
, mmu_seq
))
2211 kvm_mmu_free_some_pages(vcpu
);
2212 r
= __direct_map(vcpu
, gpa
, error_code
& PFERR_WRITE_MASK
,
2214 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2219 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2220 kvm_release_pfn_clean(pfn
);
2224 static void nonpaging_free(struct kvm_vcpu
*vcpu
)
2226 mmu_free_roots(vcpu
);
2229 static int nonpaging_init_context(struct kvm_vcpu
*vcpu
)
2231 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
2233 context
->new_cr3
= nonpaging_new_cr3
;
2234 context
->page_fault
= nonpaging_page_fault
;
2235 context
->gva_to_gpa
= nonpaging_gva_to_gpa
;
2236 context
->free
= nonpaging_free
;
2237 context
->prefetch_page
= nonpaging_prefetch_page
;
2238 context
->sync_page
= nonpaging_sync_page
;
2239 context
->invlpg
= nonpaging_invlpg
;
2240 context
->root_level
= 0;
2241 context
->shadow_root_level
= PT32E_ROOT_LEVEL
;
2242 context
->root_hpa
= INVALID_PAGE
;
2246 void kvm_mmu_flush_tlb(struct kvm_vcpu
*vcpu
)
2248 ++vcpu
->stat
.tlb_flush
;
2249 kvm_x86_ops
->tlb_flush(vcpu
);
2252 static void paging_new_cr3(struct kvm_vcpu
*vcpu
)
2254 pgprintk("%s: cr3 %lx\n", __func__
, vcpu
->arch
.cr3
);
2255 mmu_free_roots(vcpu
);
2258 static void inject_page_fault(struct kvm_vcpu
*vcpu
,
2262 kvm_inject_page_fault(vcpu
, addr
, err_code
);
2265 static void paging_free(struct kvm_vcpu
*vcpu
)
2267 nonpaging_free(vcpu
);
2270 static bool is_rsvd_bits_set(struct kvm_vcpu
*vcpu
, u64 gpte
, int level
)
2274 bit7
= (gpte
>> 7) & 1;
2275 return (gpte
& vcpu
->arch
.mmu
.rsvd_bits_mask
[bit7
][level
-1]) != 0;
2279 #include "paging_tmpl.h"
2283 #include "paging_tmpl.h"
2286 static void reset_rsvds_bits_mask(struct kvm_vcpu
*vcpu
, int level
)
2288 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
2289 int maxphyaddr
= cpuid_maxphyaddr(vcpu
);
2290 u64 exb_bit_rsvd
= 0;
2293 exb_bit_rsvd
= rsvd_bits(63, 63);
2295 case PT32_ROOT_LEVEL
:
2296 /* no rsvd bits for 2 level 4K page table entries */
2297 context
->rsvd_bits_mask
[0][1] = 0;
2298 context
->rsvd_bits_mask
[0][0] = 0;
2299 if (is_cpuid_PSE36())
2300 /* 36bits PSE 4MB page */
2301 context
->rsvd_bits_mask
[1][1] = rsvd_bits(17, 21);
2303 /* 32 bits PSE 4MB page */
2304 context
->rsvd_bits_mask
[1][1] = rsvd_bits(13, 21);
2305 context
->rsvd_bits_mask
[1][0] = context
->rsvd_bits_mask
[1][0];
2307 case PT32E_ROOT_LEVEL
:
2308 context
->rsvd_bits_mask
[0][2] =
2309 rsvd_bits(maxphyaddr
, 63) |
2310 rsvd_bits(7, 8) | rsvd_bits(1, 2); /* PDPTE */
2311 context
->rsvd_bits_mask
[0][1] = exb_bit_rsvd
|
2312 rsvd_bits(maxphyaddr
, 62); /* PDE */
2313 context
->rsvd_bits_mask
[0][0] = exb_bit_rsvd
|
2314 rsvd_bits(maxphyaddr
, 62); /* PTE */
2315 context
->rsvd_bits_mask
[1][1] = exb_bit_rsvd
|
2316 rsvd_bits(maxphyaddr
, 62) |
2317 rsvd_bits(13, 20); /* large page */
2318 context
->rsvd_bits_mask
[1][0] = context
->rsvd_bits_mask
[1][0];
2320 case PT64_ROOT_LEVEL
:
2321 context
->rsvd_bits_mask
[0][3] = exb_bit_rsvd
|
2322 rsvd_bits(maxphyaddr
, 51) | rsvd_bits(7, 8);
2323 context
->rsvd_bits_mask
[0][2] = exb_bit_rsvd
|
2324 rsvd_bits(maxphyaddr
, 51) | rsvd_bits(7, 8);
2325 context
->rsvd_bits_mask
[0][1] = exb_bit_rsvd
|
2326 rsvd_bits(maxphyaddr
, 51);
2327 context
->rsvd_bits_mask
[0][0] = exb_bit_rsvd
|
2328 rsvd_bits(maxphyaddr
, 51);
2329 context
->rsvd_bits_mask
[1][3] = context
->rsvd_bits_mask
[0][3];
2330 context
->rsvd_bits_mask
[1][2] = exb_bit_rsvd
|
2331 rsvd_bits(maxphyaddr
, 51) |
2333 context
->rsvd_bits_mask
[1][1] = exb_bit_rsvd
|
2334 rsvd_bits(maxphyaddr
, 51) |
2335 rsvd_bits(13, 20); /* large page */
2336 context
->rsvd_bits_mask
[1][0] = context
->rsvd_bits_mask
[1][0];
2341 static int paging64_init_context_common(struct kvm_vcpu
*vcpu
, int level
)
2343 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
2345 ASSERT(is_pae(vcpu
));
2346 context
->new_cr3
= paging_new_cr3
;
2347 context
->page_fault
= paging64_page_fault
;
2348 context
->gva_to_gpa
= paging64_gva_to_gpa
;
2349 context
->prefetch_page
= paging64_prefetch_page
;
2350 context
->sync_page
= paging64_sync_page
;
2351 context
->invlpg
= paging64_invlpg
;
2352 context
->free
= paging_free
;
2353 context
->root_level
= level
;
2354 context
->shadow_root_level
= level
;
2355 context
->root_hpa
= INVALID_PAGE
;
2359 static int paging64_init_context(struct kvm_vcpu
*vcpu
)
2361 reset_rsvds_bits_mask(vcpu
, PT64_ROOT_LEVEL
);
2362 return paging64_init_context_common(vcpu
, PT64_ROOT_LEVEL
);
2365 static int paging32_init_context(struct kvm_vcpu
*vcpu
)
2367 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
2369 reset_rsvds_bits_mask(vcpu
, PT32_ROOT_LEVEL
);
2370 context
->new_cr3
= paging_new_cr3
;
2371 context
->page_fault
= paging32_page_fault
;
2372 context
->gva_to_gpa
= paging32_gva_to_gpa
;
2373 context
->free
= paging_free
;
2374 context
->prefetch_page
= paging32_prefetch_page
;
2375 context
->sync_page
= paging32_sync_page
;
2376 context
->invlpg
= paging32_invlpg
;
2377 context
->root_level
= PT32_ROOT_LEVEL
;
2378 context
->shadow_root_level
= PT32E_ROOT_LEVEL
;
2379 context
->root_hpa
= INVALID_PAGE
;
2383 static int paging32E_init_context(struct kvm_vcpu
*vcpu
)
2385 reset_rsvds_bits_mask(vcpu
, PT32E_ROOT_LEVEL
);
2386 return paging64_init_context_common(vcpu
, PT32E_ROOT_LEVEL
);
2389 static int init_kvm_tdp_mmu(struct kvm_vcpu
*vcpu
)
2391 struct kvm_mmu
*context
= &vcpu
->arch
.mmu
;
2393 context
->new_cr3
= nonpaging_new_cr3
;
2394 context
->page_fault
= tdp_page_fault
;
2395 context
->free
= nonpaging_free
;
2396 context
->prefetch_page
= nonpaging_prefetch_page
;
2397 context
->sync_page
= nonpaging_sync_page
;
2398 context
->invlpg
= nonpaging_invlpg
;
2399 context
->shadow_root_level
= kvm_x86_ops
->get_tdp_level();
2400 context
->root_hpa
= INVALID_PAGE
;
2402 if (!is_paging(vcpu
)) {
2403 context
->gva_to_gpa
= nonpaging_gva_to_gpa
;
2404 context
->root_level
= 0;
2405 } else if (is_long_mode(vcpu
)) {
2406 reset_rsvds_bits_mask(vcpu
, PT64_ROOT_LEVEL
);
2407 context
->gva_to_gpa
= paging64_gva_to_gpa
;
2408 context
->root_level
= PT64_ROOT_LEVEL
;
2409 } else if (is_pae(vcpu
)) {
2410 reset_rsvds_bits_mask(vcpu
, PT32E_ROOT_LEVEL
);
2411 context
->gva_to_gpa
= paging64_gva_to_gpa
;
2412 context
->root_level
= PT32E_ROOT_LEVEL
;
2414 reset_rsvds_bits_mask(vcpu
, PT32_ROOT_LEVEL
);
2415 context
->gva_to_gpa
= paging32_gva_to_gpa
;
2416 context
->root_level
= PT32_ROOT_LEVEL
;
2422 static int init_kvm_softmmu(struct kvm_vcpu
*vcpu
)
2427 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
2429 if (!is_paging(vcpu
))
2430 r
= nonpaging_init_context(vcpu
);
2431 else if (is_long_mode(vcpu
))
2432 r
= paging64_init_context(vcpu
);
2433 else if (is_pae(vcpu
))
2434 r
= paging32E_init_context(vcpu
);
2436 r
= paging32_init_context(vcpu
);
2438 vcpu
->arch
.mmu
.base_role
.glevels
= vcpu
->arch
.mmu
.root_level
;
2443 static int init_kvm_mmu(struct kvm_vcpu
*vcpu
)
2445 vcpu
->arch
.update_pte
.pfn
= bad_pfn
;
2448 return init_kvm_tdp_mmu(vcpu
);
2450 return init_kvm_softmmu(vcpu
);
2453 static void destroy_kvm_mmu(struct kvm_vcpu
*vcpu
)
2456 if (VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
)) {
2457 vcpu
->arch
.mmu
.free(vcpu
);
2458 vcpu
->arch
.mmu
.root_hpa
= INVALID_PAGE
;
2462 int kvm_mmu_reset_context(struct kvm_vcpu
*vcpu
)
2464 destroy_kvm_mmu(vcpu
);
2465 return init_kvm_mmu(vcpu
);
2467 EXPORT_SYMBOL_GPL(kvm_mmu_reset_context
);
2469 int kvm_mmu_load(struct kvm_vcpu
*vcpu
)
2473 r
= mmu_topup_memory_caches(vcpu
);
2476 spin_lock(&vcpu
->kvm
->mmu_lock
);
2477 kvm_mmu_free_some_pages(vcpu
);
2478 r
= mmu_alloc_roots(vcpu
);
2479 mmu_sync_roots(vcpu
);
2480 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2483 /* set_cr3() should ensure TLB has been flushed */
2484 kvm_x86_ops
->set_cr3(vcpu
, vcpu
->arch
.mmu
.root_hpa
);
2488 EXPORT_SYMBOL_GPL(kvm_mmu_load
);
2490 void kvm_mmu_unload(struct kvm_vcpu
*vcpu
)
2492 mmu_free_roots(vcpu
);
2495 static void mmu_pte_write_zap_pte(struct kvm_vcpu
*vcpu
,
2496 struct kvm_mmu_page
*sp
,
2500 struct kvm_mmu_page
*child
;
2503 if (is_shadow_present_pte(pte
)) {
2504 if (is_last_spte(pte
, sp
->role
.level
))
2505 rmap_remove(vcpu
->kvm
, spte
);
2507 child
= page_header(pte
& PT64_BASE_ADDR_MASK
);
2508 mmu_page_remove_parent_pte(child
, spte
);
2511 __set_spte(spte
, shadow_trap_nonpresent_pte
);
2512 if (is_large_pte(pte
))
2513 --vcpu
->kvm
->stat
.lpages
;
2516 static void mmu_pte_write_new_pte(struct kvm_vcpu
*vcpu
,
2517 struct kvm_mmu_page
*sp
,
2521 if (sp
->role
.level
!= PT_PAGE_TABLE_LEVEL
) {
2522 ++vcpu
->kvm
->stat
.mmu_pde_zapped
;
2526 ++vcpu
->kvm
->stat
.mmu_pte_updated
;
2527 if (sp
->role
.glevels
== PT32_ROOT_LEVEL
)
2528 paging32_update_pte(vcpu
, sp
, spte
, new);
2530 paging64_update_pte(vcpu
, sp
, spte
, new);
2533 static bool need_remote_flush(u64 old
, u64
new)
2535 if (!is_shadow_present_pte(old
))
2537 if (!is_shadow_present_pte(new))
2539 if ((old
^ new) & PT64_BASE_ADDR_MASK
)
2541 old
^= PT64_NX_MASK
;
2542 new ^= PT64_NX_MASK
;
2543 return (old
& ~new & PT64_PERM_MASK
) != 0;
2546 static void mmu_pte_write_flush_tlb(struct kvm_vcpu
*vcpu
, u64 old
, u64
new)
2548 if (need_remote_flush(old
, new))
2549 kvm_flush_remote_tlbs(vcpu
->kvm
);
2551 kvm_mmu_flush_tlb(vcpu
);
2554 static bool last_updated_pte_accessed(struct kvm_vcpu
*vcpu
)
2556 u64
*spte
= vcpu
->arch
.last_pte_updated
;
2558 return !!(spte
&& (*spte
& shadow_accessed_mask
));
2561 static void mmu_guess_page_from_pte_write(struct kvm_vcpu
*vcpu
, gpa_t gpa
,
2562 const u8
*new, int bytes
)
2569 if (bytes
!= 4 && bytes
!= 8)
2573 * Assume that the pte write on a page table of the same type
2574 * as the current vcpu paging mode. This is nearly always true
2575 * (might be false while changing modes). Note it is verified later
2579 /* Handle a 32-bit guest writing two halves of a 64-bit gpte */
2580 if ((bytes
== 4) && (gpa
% 4 == 0)) {
2581 r
= kvm_read_guest(vcpu
->kvm
, gpa
& ~(u64
)7, &gpte
, 8);
2584 memcpy((void *)&gpte
+ (gpa
% 8), new, 4);
2585 } else if ((bytes
== 8) && (gpa
% 8 == 0)) {
2586 memcpy((void *)&gpte
, new, 8);
2589 if ((bytes
== 4) && (gpa
% 4 == 0))
2590 memcpy((void *)&gpte
, new, 4);
2592 if (!is_present_gpte(gpte
))
2594 gfn
= (gpte
& PT64_BASE_ADDR_MASK
) >> PAGE_SHIFT
;
2596 vcpu
->arch
.update_pte
.mmu_seq
= vcpu
->kvm
->mmu_notifier_seq
;
2598 pfn
= gfn_to_pfn(vcpu
->kvm
, gfn
);
2600 if (is_error_pfn(pfn
)) {
2601 kvm_release_pfn_clean(pfn
);
2604 vcpu
->arch
.update_pte
.gfn
= gfn
;
2605 vcpu
->arch
.update_pte
.pfn
= pfn
;
2608 static void kvm_mmu_access_page(struct kvm_vcpu
*vcpu
, gfn_t gfn
)
2610 u64
*spte
= vcpu
->arch
.last_pte_updated
;
2613 && vcpu
->arch
.last_pte_gfn
== gfn
2614 && shadow_accessed_mask
2615 && !(*spte
& shadow_accessed_mask
)
2616 && is_shadow_present_pte(*spte
))
2617 set_bit(PT_ACCESSED_SHIFT
, (unsigned long *)spte
);
2620 void kvm_mmu_pte_write(struct kvm_vcpu
*vcpu
, gpa_t gpa
,
2621 const u8
*new, int bytes
,
2622 bool guest_initiated
)
2624 gfn_t gfn
= gpa
>> PAGE_SHIFT
;
2625 struct kvm_mmu_page
*sp
;
2626 struct hlist_node
*node
, *n
;
2627 struct hlist_head
*bucket
;
2631 unsigned offset
= offset_in_page(gpa
);
2633 unsigned page_offset
;
2634 unsigned misaligned
;
2641 pgprintk("%s: gpa %llx bytes %d\n", __func__
, gpa
, bytes
);
2642 mmu_guess_page_from_pte_write(vcpu
, gpa
, new, bytes
);
2643 spin_lock(&vcpu
->kvm
->mmu_lock
);
2644 kvm_mmu_access_page(vcpu
, gfn
);
2645 kvm_mmu_free_some_pages(vcpu
);
2646 ++vcpu
->kvm
->stat
.mmu_pte_write
;
2647 kvm_mmu_audit(vcpu
, "pre pte write");
2648 if (guest_initiated
) {
2649 if (gfn
== vcpu
->arch
.last_pt_write_gfn
2650 && !last_updated_pte_accessed(vcpu
)) {
2651 ++vcpu
->arch
.last_pt_write_count
;
2652 if (vcpu
->arch
.last_pt_write_count
>= 3)
2655 vcpu
->arch
.last_pt_write_gfn
= gfn
;
2656 vcpu
->arch
.last_pt_write_count
= 1;
2657 vcpu
->arch
.last_pte_updated
= NULL
;
2660 index
= kvm_page_table_hashfn(gfn
);
2661 bucket
= &vcpu
->kvm
->arch
.mmu_page_hash
[index
];
2662 hlist_for_each_entry_safe(sp
, node
, n
, bucket
, hash_link
) {
2663 if (sp
->gfn
!= gfn
|| sp
->role
.direct
|| sp
->role
.invalid
)
2665 pte_size
= sp
->role
.glevels
== PT32_ROOT_LEVEL
? 4 : 8;
2666 misaligned
= (offset
^ (offset
+ bytes
- 1)) & ~(pte_size
- 1);
2667 misaligned
|= bytes
< 4;
2668 if (misaligned
|| flooded
) {
2670 * Misaligned accesses are too much trouble to fix
2671 * up; also, they usually indicate a page is not used
2674 * If we're seeing too many writes to a page,
2675 * it may no longer be a page table, or we may be
2676 * forking, in which case it is better to unmap the
2679 pgprintk("misaligned: gpa %llx bytes %d role %x\n",
2680 gpa
, bytes
, sp
->role
.word
);
2681 if (kvm_mmu_zap_page(vcpu
->kvm
, sp
))
2683 ++vcpu
->kvm
->stat
.mmu_flooded
;
2686 page_offset
= offset
;
2687 level
= sp
->role
.level
;
2689 if (sp
->role
.glevels
== PT32_ROOT_LEVEL
) {
2690 page_offset
<<= 1; /* 32->64 */
2692 * A 32-bit pde maps 4MB while the shadow pdes map
2693 * only 2MB. So we need to double the offset again
2694 * and zap two pdes instead of one.
2696 if (level
== PT32_ROOT_LEVEL
) {
2697 page_offset
&= ~7; /* kill rounding error */
2701 quadrant
= page_offset
>> PAGE_SHIFT
;
2702 page_offset
&= ~PAGE_MASK
;
2703 if (quadrant
!= sp
->role
.quadrant
)
2706 spte
= &sp
->spt
[page_offset
/ sizeof(*spte
)];
2707 if ((gpa
& (pte_size
- 1)) || (bytes
< pte_size
)) {
2709 r
= kvm_read_guest_atomic(vcpu
->kvm
,
2710 gpa
& ~(u64
)(pte_size
- 1),
2712 new = (const void *)&gentry
;
2718 mmu_pte_write_zap_pte(vcpu
, sp
, spte
);
2720 mmu_pte_write_new_pte(vcpu
, sp
, spte
, new);
2721 mmu_pte_write_flush_tlb(vcpu
, entry
, *spte
);
2725 kvm_mmu_audit(vcpu
, "post pte write");
2726 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2727 if (!is_error_pfn(vcpu
->arch
.update_pte
.pfn
)) {
2728 kvm_release_pfn_clean(vcpu
->arch
.update_pte
.pfn
);
2729 vcpu
->arch
.update_pte
.pfn
= bad_pfn
;
2733 int kvm_mmu_unprotect_page_virt(struct kvm_vcpu
*vcpu
, gva_t gva
)
2741 gpa
= kvm_mmu_gva_to_gpa_read(vcpu
, gva
, NULL
);
2743 spin_lock(&vcpu
->kvm
->mmu_lock
);
2744 r
= kvm_mmu_unprotect_page(vcpu
->kvm
, gpa
>> PAGE_SHIFT
);
2745 spin_unlock(&vcpu
->kvm
->mmu_lock
);
2748 EXPORT_SYMBOL_GPL(kvm_mmu_unprotect_page_virt
);
2750 void __kvm_mmu_free_some_pages(struct kvm_vcpu
*vcpu
)
2752 while (vcpu
->kvm
->arch
.n_free_mmu_pages
< KVM_REFILL_PAGES
&&
2753 !list_empty(&vcpu
->kvm
->arch
.active_mmu_pages
)) {
2754 struct kvm_mmu_page
*sp
;
2756 sp
= container_of(vcpu
->kvm
->arch
.active_mmu_pages
.prev
,
2757 struct kvm_mmu_page
, link
);
2758 kvm_mmu_zap_page(vcpu
->kvm
, sp
);
2759 ++vcpu
->kvm
->stat
.mmu_recycled
;
2763 int kvm_mmu_page_fault(struct kvm_vcpu
*vcpu
, gva_t cr2
, u32 error_code
)
2766 enum emulation_result er
;
2768 r
= vcpu
->arch
.mmu
.page_fault(vcpu
, cr2
, error_code
);
2777 r
= mmu_topup_memory_caches(vcpu
);
2781 er
= emulate_instruction(vcpu
, cr2
, error_code
, 0);
2786 case EMULATE_DO_MMIO
:
2787 ++vcpu
->stat
.mmio_exits
;
2790 vcpu
->run
->exit_reason
= KVM_EXIT_INTERNAL_ERROR
;
2791 vcpu
->run
->internal
.suberror
= KVM_INTERNAL_ERROR_EMULATION
;
2792 vcpu
->run
->internal
.ndata
= 0;
2800 EXPORT_SYMBOL_GPL(kvm_mmu_page_fault
);
2802 void kvm_mmu_invlpg(struct kvm_vcpu
*vcpu
, gva_t gva
)
2804 vcpu
->arch
.mmu
.invlpg(vcpu
, gva
);
2805 kvm_mmu_flush_tlb(vcpu
);
2806 ++vcpu
->stat
.invlpg
;
2808 EXPORT_SYMBOL_GPL(kvm_mmu_invlpg
);
2810 void kvm_enable_tdp(void)
2814 EXPORT_SYMBOL_GPL(kvm_enable_tdp
);
2816 void kvm_disable_tdp(void)
2818 tdp_enabled
= false;
2820 EXPORT_SYMBOL_GPL(kvm_disable_tdp
);
2822 static void free_mmu_pages(struct kvm_vcpu
*vcpu
)
2824 free_page((unsigned long)vcpu
->arch
.mmu
.pae_root
);
2827 static int alloc_mmu_pages(struct kvm_vcpu
*vcpu
)
2835 * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.
2836 * Therefore we need to allocate shadow page tables in the first
2837 * 4GB of memory, which happens to fit the DMA32 zone.
2839 page
= alloc_page(GFP_KERNEL
| __GFP_DMA32
);
2843 vcpu
->arch
.mmu
.pae_root
= page_address(page
);
2844 for (i
= 0; i
< 4; ++i
)
2845 vcpu
->arch
.mmu
.pae_root
[i
] = INVALID_PAGE
;
2850 int kvm_mmu_create(struct kvm_vcpu
*vcpu
)
2853 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
2855 return alloc_mmu_pages(vcpu
);
2858 int kvm_mmu_setup(struct kvm_vcpu
*vcpu
)
2861 ASSERT(!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
));
2863 return init_kvm_mmu(vcpu
);
2866 void kvm_mmu_destroy(struct kvm_vcpu
*vcpu
)
2870 destroy_kvm_mmu(vcpu
);
2871 free_mmu_pages(vcpu
);
2872 mmu_free_memory_caches(vcpu
);
2875 void kvm_mmu_slot_remove_write_access(struct kvm
*kvm
, int slot
)
2877 struct kvm_mmu_page
*sp
;
2879 list_for_each_entry(sp
, &kvm
->arch
.active_mmu_pages
, link
) {
2883 if (!test_bit(slot
, sp
->slot_bitmap
))
2887 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
)
2889 if (pt
[i
] & PT_WRITABLE_MASK
)
2890 pt
[i
] &= ~PT_WRITABLE_MASK
;
2892 kvm_flush_remote_tlbs(kvm
);
2895 void kvm_mmu_zap_all(struct kvm
*kvm
)
2897 struct kvm_mmu_page
*sp
, *node
;
2899 spin_lock(&kvm
->mmu_lock
);
2900 list_for_each_entry_safe(sp
, node
, &kvm
->arch
.active_mmu_pages
, link
)
2901 if (kvm_mmu_zap_page(kvm
, sp
))
2902 node
= container_of(kvm
->arch
.active_mmu_pages
.next
,
2903 struct kvm_mmu_page
, link
);
2904 spin_unlock(&kvm
->mmu_lock
);
2906 kvm_flush_remote_tlbs(kvm
);
2909 static void kvm_mmu_remove_one_alloc_mmu_page(struct kvm
*kvm
)
2911 struct kvm_mmu_page
*page
;
2913 page
= container_of(kvm
->arch
.active_mmu_pages
.prev
,
2914 struct kvm_mmu_page
, link
);
2915 kvm_mmu_zap_page(kvm
, page
);
2918 static int mmu_shrink(int nr_to_scan
, gfp_t gfp_mask
)
2921 struct kvm
*kvm_freed
= NULL
;
2922 int cache_count
= 0;
2924 spin_lock(&kvm_lock
);
2926 list_for_each_entry(kvm
, &vm_list
, vm_list
) {
2929 idx
= srcu_read_lock(&kvm
->srcu
);
2930 spin_lock(&kvm
->mmu_lock
);
2931 npages
= kvm
->arch
.n_alloc_mmu_pages
-
2932 kvm
->arch
.n_free_mmu_pages
;
2933 cache_count
+= npages
;
2934 if (!kvm_freed
&& nr_to_scan
> 0 && npages
> 0) {
2935 kvm_mmu_remove_one_alloc_mmu_page(kvm
);
2941 spin_unlock(&kvm
->mmu_lock
);
2942 srcu_read_unlock(&kvm
->srcu
, idx
);
2945 list_move_tail(&kvm_freed
->vm_list
, &vm_list
);
2947 spin_unlock(&kvm_lock
);
2952 static struct shrinker mmu_shrinker
= {
2953 .shrink
= mmu_shrink
,
2954 .seeks
= DEFAULT_SEEKS
* 10,
2957 static void mmu_destroy_caches(void)
2959 if (pte_chain_cache
)
2960 kmem_cache_destroy(pte_chain_cache
);
2961 if (rmap_desc_cache
)
2962 kmem_cache_destroy(rmap_desc_cache
);
2963 if (mmu_page_header_cache
)
2964 kmem_cache_destroy(mmu_page_header_cache
);
2967 void kvm_mmu_module_exit(void)
2969 mmu_destroy_caches();
2970 unregister_shrinker(&mmu_shrinker
);
2973 int kvm_mmu_module_init(void)
2975 pte_chain_cache
= kmem_cache_create("kvm_pte_chain",
2976 sizeof(struct kvm_pte_chain
),
2978 if (!pte_chain_cache
)
2980 rmap_desc_cache
= kmem_cache_create("kvm_rmap_desc",
2981 sizeof(struct kvm_rmap_desc
),
2983 if (!rmap_desc_cache
)
2986 mmu_page_header_cache
= kmem_cache_create("kvm_mmu_page_header",
2987 sizeof(struct kvm_mmu_page
),
2989 if (!mmu_page_header_cache
)
2992 register_shrinker(&mmu_shrinker
);
2997 mmu_destroy_caches();
3002 * Caculate mmu pages needed for kvm.
3004 unsigned int kvm_mmu_calculate_mmu_pages(struct kvm
*kvm
)
3007 unsigned int nr_mmu_pages
;
3008 unsigned int nr_pages
= 0;
3009 struct kvm_memslots
*slots
;
3011 slots
= rcu_dereference(kvm
->memslots
);
3012 for (i
= 0; i
< slots
->nmemslots
; i
++)
3013 nr_pages
+= slots
->memslots
[i
].npages
;
3015 nr_mmu_pages
= nr_pages
* KVM_PERMILLE_MMU_PAGES
/ 1000;
3016 nr_mmu_pages
= max(nr_mmu_pages
,
3017 (unsigned int) KVM_MIN_ALLOC_MMU_PAGES
);
3019 return nr_mmu_pages
;
3022 static void *pv_mmu_peek_buffer(struct kvm_pv_mmu_op_buffer
*buffer
,
3025 if (len
> buffer
->len
)
3030 static void *pv_mmu_read_buffer(struct kvm_pv_mmu_op_buffer
*buffer
,
3035 ret
= pv_mmu_peek_buffer(buffer
, len
);
3040 buffer
->processed
+= len
;
3044 static int kvm_pv_mmu_write(struct kvm_vcpu
*vcpu
,
3045 gpa_t addr
, gpa_t value
)
3050 if (!is_long_mode(vcpu
) && !is_pae(vcpu
))
3053 r
= mmu_topup_memory_caches(vcpu
);
3057 if (!emulator_write_phys(vcpu
, addr
, &value
, bytes
))
3063 static int kvm_pv_mmu_flush_tlb(struct kvm_vcpu
*vcpu
)
3065 kvm_set_cr3(vcpu
, vcpu
->arch
.cr3
);
3069 static int kvm_pv_mmu_release_pt(struct kvm_vcpu
*vcpu
, gpa_t addr
)
3071 spin_lock(&vcpu
->kvm
->mmu_lock
);
3072 mmu_unshadow(vcpu
->kvm
, addr
>> PAGE_SHIFT
);
3073 spin_unlock(&vcpu
->kvm
->mmu_lock
);
3077 static int kvm_pv_mmu_op_one(struct kvm_vcpu
*vcpu
,
3078 struct kvm_pv_mmu_op_buffer
*buffer
)
3080 struct kvm_mmu_op_header
*header
;
3082 header
= pv_mmu_peek_buffer(buffer
, sizeof *header
);
3085 switch (header
->op
) {
3086 case KVM_MMU_OP_WRITE_PTE
: {
3087 struct kvm_mmu_op_write_pte
*wpte
;
3089 wpte
= pv_mmu_read_buffer(buffer
, sizeof *wpte
);
3092 return kvm_pv_mmu_write(vcpu
, wpte
->pte_phys
,
3095 case KVM_MMU_OP_FLUSH_TLB
: {
3096 struct kvm_mmu_op_flush_tlb
*ftlb
;
3098 ftlb
= pv_mmu_read_buffer(buffer
, sizeof *ftlb
);
3101 return kvm_pv_mmu_flush_tlb(vcpu
);
3103 case KVM_MMU_OP_RELEASE_PT
: {
3104 struct kvm_mmu_op_release_pt
*rpt
;
3106 rpt
= pv_mmu_read_buffer(buffer
, sizeof *rpt
);
3109 return kvm_pv_mmu_release_pt(vcpu
, rpt
->pt_phys
);
3115 int kvm_pv_mmu_op(struct kvm_vcpu
*vcpu
, unsigned long bytes
,
3116 gpa_t addr
, unsigned long *ret
)
3119 struct kvm_pv_mmu_op_buffer
*buffer
= &vcpu
->arch
.mmu_op_buffer
;
3121 buffer
->ptr
= buffer
->buf
;
3122 buffer
->len
= min_t(unsigned long, bytes
, sizeof buffer
->buf
);
3123 buffer
->processed
= 0;
3125 r
= kvm_read_guest(vcpu
->kvm
, addr
, buffer
->buf
, buffer
->len
);
3129 while (buffer
->len
) {
3130 r
= kvm_pv_mmu_op_one(vcpu
, buffer
);
3139 *ret
= buffer
->processed
;
3143 int kvm_mmu_get_spte_hierarchy(struct kvm_vcpu
*vcpu
, u64 addr
, u64 sptes
[4])
3145 struct kvm_shadow_walk_iterator iterator
;
3148 spin_lock(&vcpu
->kvm
->mmu_lock
);
3149 for_each_shadow_entry(vcpu
, addr
, iterator
) {
3150 sptes
[iterator
.level
-1] = *iterator
.sptep
;
3152 if (!is_shadow_present_pte(*iterator
.sptep
))
3155 spin_unlock(&vcpu
->kvm
->mmu_lock
);
3159 EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy
);
3163 static const char *audit_msg
;
3165 static gva_t
canonicalize(gva_t gva
)
3167 #ifdef CONFIG_X86_64
3168 gva
= (long long)(gva
<< 16) >> 16;
3174 typedef void (*inspect_spte_fn
) (struct kvm
*kvm
, struct kvm_mmu_page
*sp
,
3177 static void __mmu_spte_walk(struct kvm
*kvm
, struct kvm_mmu_page
*sp
,
3182 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
3183 u64 ent
= sp
->spt
[i
];
3185 if (is_shadow_present_pte(ent
)) {
3186 if (!is_last_spte(ent
, sp
->role
.level
)) {
3187 struct kvm_mmu_page
*child
;
3188 child
= page_header(ent
& PT64_BASE_ADDR_MASK
);
3189 __mmu_spte_walk(kvm
, child
, fn
);
3191 fn(kvm
, sp
, &sp
->spt
[i
]);
3196 static void mmu_spte_walk(struct kvm_vcpu
*vcpu
, inspect_spte_fn fn
)
3199 struct kvm_mmu_page
*sp
;
3201 if (!VALID_PAGE(vcpu
->arch
.mmu
.root_hpa
))
3203 if (vcpu
->arch
.mmu
.shadow_root_level
== PT64_ROOT_LEVEL
) {
3204 hpa_t root
= vcpu
->arch
.mmu
.root_hpa
;
3205 sp
= page_header(root
);
3206 __mmu_spte_walk(vcpu
->kvm
, sp
, fn
);
3209 for (i
= 0; i
< 4; ++i
) {
3210 hpa_t root
= vcpu
->arch
.mmu
.pae_root
[i
];
3212 if (root
&& VALID_PAGE(root
)) {
3213 root
&= PT64_BASE_ADDR_MASK
;
3214 sp
= page_header(root
);
3215 __mmu_spte_walk(vcpu
->kvm
, sp
, fn
);
3221 static void audit_mappings_page(struct kvm_vcpu
*vcpu
, u64 page_pte
,
3222 gva_t va
, int level
)
3224 u64
*pt
= __va(page_pte
& PT64_BASE_ADDR_MASK
);
3226 gva_t va_delta
= 1ul << (PAGE_SHIFT
+ 9 * (level
- 1));
3228 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
, va
+= va_delta
) {
3231 if (ent
== shadow_trap_nonpresent_pte
)
3234 va
= canonicalize(va
);
3235 if (is_shadow_present_pte(ent
) && !is_last_spte(ent
, level
))
3236 audit_mappings_page(vcpu
, ent
, va
, level
- 1);
3238 gpa_t gpa
= kvm_mmu_gva_to_gpa_read(vcpu
, va
, NULL
);
3239 gfn_t gfn
= gpa
>> PAGE_SHIFT
;
3240 pfn_t pfn
= gfn_to_pfn(vcpu
->kvm
, gfn
);
3241 hpa_t hpa
= (hpa_t
)pfn
<< PAGE_SHIFT
;
3243 if (is_error_pfn(pfn
)) {
3244 kvm_release_pfn_clean(pfn
);
3248 if (is_shadow_present_pte(ent
)
3249 && (ent
& PT64_BASE_ADDR_MASK
) != hpa
)
3250 printk(KERN_ERR
"xx audit error: (%s) levels %d"
3251 " gva %lx gpa %llx hpa %llx ent %llx %d\n",
3252 audit_msg
, vcpu
->arch
.mmu
.root_level
,
3254 is_shadow_present_pte(ent
));
3255 else if (ent
== shadow_notrap_nonpresent_pte
3256 && !is_error_hpa(hpa
))
3257 printk(KERN_ERR
"audit: (%s) notrap shadow,"
3258 " valid guest gva %lx\n", audit_msg
, va
);
3259 kvm_release_pfn_clean(pfn
);
3265 static void audit_mappings(struct kvm_vcpu
*vcpu
)
3269 if (vcpu
->arch
.mmu
.root_level
== 4)
3270 audit_mappings_page(vcpu
, vcpu
->arch
.mmu
.root_hpa
, 0, 4);
3272 for (i
= 0; i
< 4; ++i
)
3273 if (vcpu
->arch
.mmu
.pae_root
[i
] & PT_PRESENT_MASK
)
3274 audit_mappings_page(vcpu
,
3275 vcpu
->arch
.mmu
.pae_root
[i
],
3280 static int count_rmaps(struct kvm_vcpu
*vcpu
)
3285 idx
= srcu_read_lock(&kvm
->srcu
);
3286 slots
= rcu_dereference(kvm
->memslots
);
3287 for (i
= 0; i
< KVM_MEMORY_SLOTS
; ++i
) {
3288 struct kvm_memory_slot
*m
= &slots
->memslots
[i
];
3289 struct kvm_rmap_desc
*d
;
3291 for (j
= 0; j
< m
->npages
; ++j
) {
3292 unsigned long *rmapp
= &m
->rmap
[j
];
3296 if (!(*rmapp
& 1)) {
3300 d
= (struct kvm_rmap_desc
*)(*rmapp
& ~1ul);
3302 for (k
= 0; k
< RMAP_EXT
; ++k
)
3311 srcu_read_unlock(&kvm
->srcu
, idx
);
3315 void inspect_spte_has_rmap(struct kvm
*kvm
, struct kvm_mmu_page
*sp
, u64
*sptep
)
3317 unsigned long *rmapp
;
3318 struct kvm_mmu_page
*rev_sp
;
3321 if (*sptep
& PT_WRITABLE_MASK
) {
3322 rev_sp
= page_header(__pa(sptep
));
3323 gfn
= rev_sp
->gfns
[sptep
- rev_sp
->spt
];
3325 if (!gfn_to_memslot(kvm
, gfn
)) {
3326 if (!printk_ratelimit())
3328 printk(KERN_ERR
"%s: no memslot for gfn %ld\n",
3330 printk(KERN_ERR
"%s: index %ld of sp (gfn=%lx)\n",
3331 audit_msg
, sptep
- rev_sp
->spt
,
3337 rmapp
= gfn_to_rmap(kvm
, rev_sp
->gfns
[sptep
- rev_sp
->spt
],
3338 is_large_pte(*sptep
));
3340 if (!printk_ratelimit())
3342 printk(KERN_ERR
"%s: no rmap for writable spte %llx\n",
3350 void audit_writable_sptes_have_rmaps(struct kvm_vcpu
*vcpu
)
3352 mmu_spte_walk(vcpu
, inspect_spte_has_rmap
);
3355 static void check_writable_mappings_rmap(struct kvm_vcpu
*vcpu
)
3357 struct kvm_mmu_page
*sp
;
3360 list_for_each_entry(sp
, &vcpu
->kvm
->arch
.active_mmu_pages
, link
) {
3363 if (sp
->role
.level
!= PT_PAGE_TABLE_LEVEL
)
3366 for (i
= 0; i
< PT64_ENT_PER_PAGE
; ++i
) {
3369 if (!(ent
& PT_PRESENT_MASK
))
3371 if (!(ent
& PT_WRITABLE_MASK
))
3373 inspect_spte_has_rmap(vcpu
->kvm
, sp
, &pt
[i
]);
3379 static void audit_rmap(struct kvm_vcpu
*vcpu
)
3381 check_writable_mappings_rmap(vcpu
);
3385 static void audit_write_protection(struct kvm_vcpu
*vcpu
)
3387 struct kvm_mmu_page
*sp
;
3388 struct kvm_memory_slot
*slot
;
3389 unsigned long *rmapp
;
3393 list_for_each_entry(sp
, &vcpu
->kvm
->arch
.active_mmu_pages
, link
) {
3394 if (sp
->role
.direct
)
3399 gfn
= unalias_gfn(vcpu
->kvm
, sp
->gfn
);
3400 slot
= gfn_to_memslot_unaliased(vcpu
->kvm
, sp
->gfn
);
3401 rmapp
= &slot
->rmap
[gfn
- slot
->base_gfn
];
3403 spte
= rmap_next(vcpu
->kvm
, rmapp
, NULL
);
3405 if (*spte
& PT_WRITABLE_MASK
)
3406 printk(KERN_ERR
"%s: (%s) shadow page has "
3407 "writable mappings: gfn %lx role %x\n",
3408 __func__
, audit_msg
, sp
->gfn
,
3410 spte
= rmap_next(vcpu
->kvm
, rmapp
, spte
);
3415 static void kvm_mmu_audit(struct kvm_vcpu
*vcpu
, const char *msg
)
3422 audit_write_protection(vcpu
);
3423 if (strcmp("pre pte write", audit_msg
) != 0)
3424 audit_mappings(vcpu
);
3425 audit_writable_sptes_have_rmaps(vcpu
);